Method for classifying a digital image

0001The present invention relates to the processing of digital images and, more particularly, relates to a method for classifying a digital image.

0002Digital images are currently being used in numerous applications, for example in new generation acquisition devices such as photo cameras and digital still cameras (DSC). Moreover, digital images are being ever more extensively employed in applications in the field of artificial vision and in applications in the field of medical diagnostics.

0003In these and other contexts image classification plays a fundamental role as integrating part of the processing procedures, because it makes it possible not only to select the most appropriate algorithms on the basis of the type of image that has to be processed, but also to render the processing procedures independent of interaction with the user.

0004In the field of artificial vision, for example, some techniques for converting two-dimensional images into three-dimensional images are sensitively bound up with the type of image to be converted.

0005In some of these conversion techniques, in particular, there is typically generated a depth map relating to the 2D image to be converted, and on the basis of this map there is subsequently created a stereoscopic pair necessary for the reconstruction of the 3D output image. In the generation of this depth map the classification of the images acts as "guide" element within the system, because it makes it possible to select and therefore utilize the most appropriate techniques on the basis of the image that is to be processed.

0006Image classification also plays a very important part in the processing procedures performed within the digital image acquisition devices. In this context the classification is utilized, for example, to optimize the processing as regards improvement of the quality of an acquired digital image or to optimize the compression-encoding operations.

0007Though widely used, the known classification techniques are associated with limits both as regards the precision of the results and from the point of view of computational complexity.

0008The present invention therefore sets out to make available a robust classification method that will be capable of providing better performances as compared with the methods of the known techniques.

0009This aim is attained by means of a classification method as described in the first claim attached hereto.

0010Preferred embodiments of the method of the invention are defined in the dependent claims.

0011Other objects of the present invention are constituted by an image acquisition device as described in Claim 26, a computer program as described in Claim 27 and a memorization support as described in Claim 28.

0012Further characteristics and advantages of the invention will be more readily understood from the following detailed description of its preferred embodiments, which is given by way of example and is not be understood as limitative in any way, of which: <ul id="ul0001" list-style="dash" compact="compact"><li>Figure 1 schematically shows a block diagram relating to a classification method in accordance with the present invention,</li><li>Figure 2 shows the pattern in which the filter elements are arranged in a sensor with an optical Bayer filter,</li><li>Figure 3 shows an input image of the segmentation block Img_Seg and a corresponding segmented image produced as output from that block,</li><li>Figure 4 shows the simplified flow chart of a preferred embodiment of the operation of assigning a grey level I_r(x,y) to a pixel p(x,y) of the segmented and filtered image F_Img_s,</li><li>Figure 5 shows an output image of the colour reconstruction block and the corresponding output images of, respectively, the segmentation block and the block that divides the image into regions,</li><li>Figure 6 shows the class decision block in greater detail by means of a flow chart, and</li><li>Figure 7 shows the scanned columns in an image divided into regions in the information acquisition operation f₁ shown in the diagram of Figure 6.</li></ul>

0013Figure 1 schematically represents a block diagram of a classification method Img_Class in accordance with a preferred embodiment of the present invention

0014As shown in Figure 1, the classification method Img_Class preferably comprises a colour reconstruction block Col_Rec, a segmentation block Img_Seg, a block Reg_Detect for dividing the image into regions and a class decision block.

0015The color reconstruction block Col_Rec is such as to carry out a processing phase of chromatic reconstruction, by means of which, starting from a digital input image Img_CFA in CFA (Color Filter Array) format, i.e. including pixels that are each associated with a respective digital value (represented, for example, by 8 bits), one obtains a complete digital output image I_RGB, i.e. including pixels that are each associated with a respective triad of digital values.

0016The input digital image Img_CFA of the colour reconstruction block Col_Rec is preferably a digital image acquired with a sensor of the CCD (Charge Coupled Device) or CMOS (Complementary Metal Oxide Semiconductor) type comprising an optical CFA filter, a Bayer matrix filter for example.

0017As is well known to persons skilled in the art, only a single photosensitive cell is available in a sensor with a CFA filter for acquiring a pixel of the image. The sensor is covered by an optical filter comprising a matrix (a Bayer matrix, for example) of filter elements, each of which is associated with a photosensitive cell. Each filter element transmits to the photoelectric cell associated with it light radiation corresponding to the wavelength of only red light, only green light or only blue light (absorbing only a minimal part of it) and thus reveals only one component of each pixel.

0018The layout pattern BP of the filter elements in a Bayer filter is represented in Figure 2, where R, G and B indicate, respectively, the red, green and blue filter elements.

0019The chromatic reconstruction phase is then used to recover at least partially the chromatic information present in the real scene. Preferably, the output image Img_RGB of the reconstruction block is a digital image in RGB (Red Green Blue) format, i.e. an image in which three digital values corresponding, respectively, to the chromatic components R (Red), G (Green) and B (Blue) are associated with each pixel.

0020In the chromatic reconstruction phase in order to obtain an output image Img_RGB in RGB format from an input image Img_CFA in Bayer CFA format there is carried out a colour interpolation operation that is known to persons skilled in the art and will not therefore be further described.

0021In a particularly advantageous variant embodiment, the chromatic reconstruction phase comprises an operation of sub-sampling the input image Img_CFA in CFA format.

0022Preferably, in the sub-sampling operation the digital input image Img_CFA is segmented into small adjacent and not overlapping blocks of pixels, for example into blocks of size 2x2. Starting from each of these small blocks, there is generated a respective coloured pixel in the output image Img_RGB in RGB format, of which the R and B components correspond to the digital values of the pixels on the secondary diagonal of the block and of which the G component is equal to the mean of the digital values of the pixels on the principal diagonal of the block.

0023In this way, starting from each block of size 2x2 of the input image Img_CFA in Bayer CFA format, the colour reconstruction phase produces a single pixel of the output image Img_RGB in RGB format.

0024The resolution of the output image is therefore reduced with respect to the resolution of the input image: in this specific example it is equal to a quarter of the resolution of the input image.

0025Referring to the block pattern represented in Figure 1, the segmentation block Img_Seg that receives the input image Img_RGB is such as to produce a segmented output image Img_s, i.e. divided into areas that are substantially homogeneous from the point of view of chromatic appearance.

0026In greater detail, in the segmentation block Img_Seg there is carried out a processing phase that makes it possible to identify the significant chromatic characteristics present in an image. Preferably, this processing phase is such as to operate in accordance with a technique known by the name of "mean shift algorithm", a simple and non-parametric procedure for estimating the density gradients. Advantageously, this technique is found to be particularly efficient from the computational point of view. Nevertheless, other types of segmentation may also be utilized.

0027The mean shift algorithm technique is described, for example, in the document: "Robust Analysis of Feature Spaces: Color Image Segmentation" (D. Comaniciu, P. Meer, Department of Electrical and Computer Engineering, Rutgers University, Piscataway, U.S.A).

0028In accordance with this technique, the processing phase carried out in the segmentation block Img_Seg preferably includes a preliminary phase of transforming the digital values of the pixels of the image in RGB format from the chromatic RGB space to the Luv space.

0029The passage from the RGB space to the Luv space calls for a non-linear transformation.

0030As explained in the aforementioned article by Comaniciu and Meer, this transformation operation makes it possible for the processing operations carried out in the segmentation block to operate advantageously on a sufficiently isotropic space of the components.

0031The processing phase carried out in the segmentation block Img_Seg also includes a successive operation of grouping the pixels in an area on the basis of their resemblance, that is to say, the homogeneity of the pixels transformed in this manner.

0032Lastly, a successive operation included in the processing phase is such as to modify the pixels belonging to one and the same area in such a manner as to associate with them digital values that are substantially representative of the predominant colours of the area, thereby obtaining a segmented image Img_s that, as compared with the image Img_RGB present at the input of the segmentation block Img_Seg, comprises a reduced range of colours.

0033Referring to the aforementioned publication by Comaniciu and Meer, the processing phase performed in the segmentation block Img_Seg operates in accordance with the mean shift algorithm technique having a low segmentation resolution (under-segmentation).

0034According to this technique, the pixels are grouped together in areas by means of homogeneity criteria having an ample margin of tolerance and thus obtaining a segmented image in which the pixels have associated with them only the most significant (i.e. predominant) colours present in the input image Img_RGB of the segmentation block.

0035It should be borne in mind that the number of colours present in a generic image depends on the quantity of detail present and is typically comprised between 35,000 and 40,000.

0036In a particularly preferred embodiment the segmentation block Img_Seg produces an output image Img_s in RGB format comprising at the most about ten colours and is therefore more homogeneous and compact than the input image Img_RGB.

0037In Figure 3 Img_RGB is used to indicate an input image of the segmentation block Img_Seg, while Img_s indicates a corresponding segmented output image produced by means of an under-segmentation processing phase. Although the images in Figure 3 are represented in grey scale, one can nevertheless observe the effect of the colour reduction from the input image (in this example with 37,656 colours) to the output image (in this example with 4 colours).

0038Advantageously, it has been found that the segmentation, and particularly the "under-segmentation", makes it possible to simplify the work of the subsequent block and, more particularly, makes it possible to reduce the computational cost and at the same time to increase the robustness of the classification method in accordance with the present invention.

0039Coming back to the block diagram shown in Figure 1, the region detection block Reg_Detect receives the segmented input image Img_s and carries out a division phase that makes it possible to divide this image, pixel by pixel, into one or more regions forming part of a set of predefined regions that differ from each other on the basis of the type of content, producing an output image Img_Reg divided into one or more regions.

0040Preferably, the predefined set of regions comprises regions of different typologies that differ from each other on account of the type of content and are identified as: "Sky", "Farthest mountain", "Far mountain", "Near mountain", "Land" and "Other". Nevertheless, regions of different types can also be taken into consideration.

0041It should be noted that the regions listed above are typically present in panoramic images, landscapes or, in any case, outdoor photographs.

0042In a particularly preferred embodiment, the division into regions is effected by establishing whether or not a pixel of the digital image belongs to a respective region on the basis of an operation analyzing the parameters of that pixel. The analysis operation is carried out by verifying whether one or more parameters of the pixel satisfy predetermined conditions and/or logico-mathematical relationships that characterize membership of the respective region. The conditions and the logico-mathematical conditions that characterize membership of a predefined region are derived empirically from a statistical observation of images that depict real scenes.

0043Preferably, the analyzed parameters will correspond to digital values representative of at least one of the following components of the pixel: red chromatic component, green chromatic component, blue chromatic component, hue and intensity. Other parameters similar to those just cited may however be analyzed.

0044After the analysis operation, the division phase also includes an operation of assigning to the analyzed pixel, i.e. associating it with a data item that identifies the respective region to which it belongs.

0045In an advantageous embodiment the identification data item comprises a digital value, in practice a grey level.

0046For example, with pixels belonging to predefined regions that on the basis of their content can be considered to be nearer to an observation point of the image there are associated respective darker grey levels than those of pixels belonging to more distant regions. Obviously, a choice of the opposite type is wholly equivalent.

0047It should be noted that the set of digital values determined by the operation of assigning a digital value to the pixels of the digital image to be divided corresponds to a grey-level digital image divided into regions.

0048In a particular example, on the assumption that the pixels comprise 8-bit digital values, the digital values (i.e. the grey levels) to be assigned to the pixels on the basis of the region to which they belong are, for example, as shown in the following table: <tables id="tabl0001" num="0001"><table frame="all"><tgroup cols="2" colsep="1" rowsep="1"><colspec colnum="1" colname="col1" colwidth="78.75mm" /><colspec colnum="2" colname="col2" colwidth="78.75mm" /><thead valign="top"><row><entry namest="col1" nameend="col1" align="center"><b>REGION</b></entry><entry namest="col2" nameend="col2" align="center"><b>VALUE</b></entry></row></thead><tbody valign="top"><row><entry namest="col1" nameend="col1" align="left">Sky</entry><entry namest="col2" nameend="col2" align="left">0</entry></row><row><entry namest="col1" nameend="col1" align="left">Farthest mountain</entry><entry namest="col2" nameend="col2" align="left">50</entry></row><row><entry namest="col1" nameend="col1" align="left">Far mountain</entry><entry namest="col2" nameend="col2" align="left">80</entry></row><row><entry namest="col1" nameend="col1" align="left">Near mountain</entry><entry namest="col2" nameend="col2" align="left">100</entry></row><row><entry namest="col1" nameend="col1" align="left">Land</entry><entry namest="col2" nameend="col2" align="left">200</entry></row><row rowsep="1"><entry namest="col1" nameend="col1" align="left">Other</entry><entry namest="col2" nameend="col2" align="left">255</entry></row></tbody></tgroup></table></tables>

0049In a particularly preferred embodiment, the division into regions is effected by analyzing the image pixel by pixel: for every pixel of the segmented image there are extracted the values of the components R,G,B and the components H (Hue) and I (Intensity) and on the basis of these values it is decided whether or not a pixel belongs to one of the regions listed above is then established on the basis of these values.

0050The characterization of the regions is effected by means of heuristic rules and, more particularly, by conditioning rules: the conditions are represented by intervals and logico-mathematical relationships between the values R, G, B, H and I, while the corresponding actions are represented by the association with the analyzed pixels of a grey-level value.

0051In a particularly preferred embodiment, the phase of dividing the segmented image performed in the division block Reg_Detect comprises the operations of: <ul id="ul0002" list-style="dash" compact="compact"><li>application of a median filter, preferably 5x5, to the segmented image Img_s for obtaining a filtered segmented image F_Img_s;</li><li>extraction of the digital values associated with the components RGB and HSI of the filtered segmented image F_Img_s;</li><li>association with each pixel of the filtered segmented image F_Img_s of a grey level; associating a grey level to each pixel of the segmented image one obtains in practice a grey-level image temp_Img_Reg;</li><li>application of a median filter, preferably 5x5, to the grey-level image temp_Img_Reg to obtain a final grey-level image Img_Reg.</li></ul>

0052Advantageously, the use of median filters makes it possible to reduce the noise component present in both the input image and the output image of the block for the division into regions.

0053Figure 4 shows the flow chart of a preferred embodiment of the operation of assigning a grey-level I_r(x,y) to a pixel p(x,y) of the filtered segmented image F_Img_s to obtain a corresponding pixel I_r(x,y) of the grey-level image temp_Img_Reg.

0054A check is carried out in a first control phase Ck1 for establishing whether or not the pixel p(x,y) belongs to a region of the type "Near mountain" indicated in the figure by the reference "r1".

0055If on the basis of this check the pixel p(x,y) is found to belong to this region, a first phase of association Set1 assigns a grey level having a digital value equal to 100 to the pixel I_r(x,y) of the grey-level image temp_Img_Reg. In that case the association operation terminates and a new pixel is taken into consideration. Otherwise, a second control phase Ck2 is carried out to establish whether or not the pixel p(x,y) belongs to a region of the type "Far mountain", indicated in the figure by the reference "r2".

0056If on the basis of this check the pixel p(x,y) is found to belong to this region, a second phase of association Set2 assigns a grey level having a digital value equal to 80 to the pixel I_r(x,y) of the grey-level image temp_Img_Reg. In that case the association operation terminates and a new pixel is taken into consideration.

0057Otherwise, a third control phase Ck3 is carried out for establishing whether or not the pixel p(x,y) belongs to a region of the type "Far mountain", indicated in the figure by the reference "r3".

0058If on the basis of this check the pixel p(x,y) is found to belong to this region, a third phase of association Set3 assigns a grey level having a digital value equal to 50 to the pixel I_r(x,y) of the grey-level image temp_Img_Reg. In that case the association operation terminates and a new pixel is taken into consideration.

0059In a preferred embodiment the conditions that make it possible to establish whether or not the pixel p(x,y) belongs to a region of the type "Near mountain" ("r1") are expressed by the following Boolean condition: B(x,y)≥G(x,y) AND I(x,y)≤(mean_I-15) AND B(x,y)≤(G(x,y)+50) AND I(x,y)≤90<maths id="math0001" num=""><img file="EP1507232A2_D0001.tif" /></maths><maths id="math0002" num=""><img file="EP1507232A2_D0002.tif" /></maths>

0060Likewise in a preferred embodiment, the conditions that make it possible to establish whether or not the pixel p(x,y) belongs to a region of the type "Far mountain" ("r2") are expressed by the following Boolean condition: B(x,y)≥G(x,y) AND I(x,y)≤(mean_I-15) AND B(x,y)≤(G(x,y)+50) AND I(x,y) ≤130.

0061While the conditions for establishing whether or not the pixel p(x,y) belongs to a region of the type "Farthest mountain" ("r3") are expressed by the following Boolean condition: B(x,y)≥G(x,y) AND I(x,y)≤(mean_I-15) AND B(x,y)≤(G(x,y)+50) AND I(x,y)>130.

0062In general, the condition that is common to the greater part of the pixels that represent mountains is the one of having the component B greater than the component G, but limited by G+50.

0063The quantity mean_I constitutes an estimate of the mean value of the luminous intensity of the pixels that should belong to a region of the type "Sky". It is calculated as the mean of the luminous intensity of the pixels of the first ten rows of the image having the component B greater than the component G. The check of the value of the intensity I with respect to mean_I-15 makes it possible to avoid that pixels belonging to a region of the type "Sky" are mistaken for a pixel of a region of the type "Mountain".

0064In particular, the differentiation between the three types of "Mountain" is effected by exploiting the concept of atmospheric perspective introduced for the first time by Leonardo in the famous painting of the Mona Lisa. In fact, he had noted that the most distant mountains, given that they had the same chromatic hue as their nearest counterparts, appeared to be clearer (and therefore had a greater luminous intensity). This phenomenon can be explained by considering that the quantity of atmospheric dust that becomes interposed between an observer and an object increases with the distance of the object from the observer.

0065In this manner the more distant objects have colours nearer to white and therefore have a luminous intensity value greater than the nearer objects.

0066The "Farthest mountain" region is therefore characterized by a luminous intensity value I greater than 130, "Far mountain" has a luminous intensity value I comprised between 90 and 130, while "Near mountain" representing the nearest mountain is characterized by a luminous intensity value I of less than 90.

0067Coming back to the flow chart of Figure 4, if on the basis of the checks carried out in the phases Ck1, Ck2 and Ck3 the pixel p(x,y) is found not to belong to one of the regions of the "Mountain" type, a fourth control phase Ck4 is undertaken to ascertain whether or not the pixel p(x,y) belongs to a region of the "Land" type indicated in the figure by the reference "r4".

0068If on the basis of this check the pixel p(x,y) is found to belong to this region, a fourth phase of association Set4 assigns a grey level having a digital value equal to 200 to the pixel I_r(x,y) of the grey-level image temp_Img_Reg. In that case the association operation terminates and a new pixel is taken into consideration.

0069In a preferred embodiment the conditions for checking whether or not the pixel p(x,y) belongs to a region of the "Land" type ("r4") are expressed by the following Boolean conditions: <ul id="ul0003" list-style="bullet" compact="compact"><li>|R(x,y)-G(x,y)|≤10 AND R(x,y)≥(B(x,y)+55)</li><li>|R (x, y) -G (x, y)|≤30 AND |R (x, y) -B (x, y) |≤30 AND 40≤I(x,y)≤80</li><li>R(x,y)>(B(x,y)+45) AND R(x,y)≥G(x,y)</li><li>R(x,y)≥G(x,y) AND G(x,y)≥B(x,y) AND 25≤H(x,y)≤50</li><li>G(x,y)≥(R(x,y)+10) G(x,y)≥(B(x,y)+15) AND I(x,y)≥75</li><li>|R(x,y)-G(x,y)|≤12 AND G(x,y)>B(x,y) AND R(x,y)>B(x,y).</li></ul>

0070The pixel is deemed to belong to a region of the "Land" type if one of the aforesaid conditions is found to be satisfied.

0071These conditions describe the family of lands that can be present in an image of the "landscape type". The first condition describes the colours typical of a beach. The next three conditions represent colours of lands with chromatic gradations tending towards dark brown and red. The penultimate condition describes the colours typical of green meadows characterized by having the component G greater than the others and a rather high luminous intensity (greater than 75). The last condition has been inserted to represent the cultivated fields (wheat fields, for example).

0072If none of the aforementioned six conditions is satisfied, a fifth control phase Ck5 is carried to ascertain whether or not the pixel p(x,y) belongs to a region of the "Sky" type indicated in the figure by the reference "r5".

0073If on the basis of this check the pixel p(x,y) is found to belong to this region r5, a fifth phase of association Set5 assigns a grey level having a digital value equal to 0 to the pixel I_r(x,y) of the grey-level image temp_Img_Reg. In that case the association operation terminates and a new pixel is taken into consideration.

0074In a preferred embodiment the conditions for checking whether or not the pixel p(x,y) belongs to a region of the "Sky" type ("r5") are expressed by the following Boolean conditions: <ul id="ul0004" list-style="bullet" compact="compact"><li>B(x,y)≥(G(x,y)+30) AND G(x,y)≥R(x,y)</li><li>B(x,y)≥G(x,y) AND G(x,y)≥R(x,y)</li><li>B(x,y)≥R(x,y) AND R(x,y)≥G(x,y)</li><li>I(x,y)≥200 AND |B(x,y)-G(x,y)|≤30 AND |R(x,y)-G(x,y)|≤30</li></ul>

0075The pixel is deemed to belong to a region of the "Sky" type if one of the aforesaid four conditions is found to be satisfied.

0076The first three conditions represent three variations of blue skies in which the predominant component is B(x,y). They describe the most common types of sky. The last condition represents a range of cloudy skies characterized by a grey colour. Given the nature of the chromatic gradations of grey, there is no need for the presence of a component greater than the others. Furthermore, there is present a check of the value I to avoid that pixels that really belong to one of the regions of the "Mountain" type are labelled as "Sky".

0077The choice of the threshold value equal to 200 derives from the observation of the value I of cloudy skies present in the images. In fact, they are characterized by being particularly luminous. If none of the aforesaid four conditions is satisfied, the pixel p(x,y) is automatically assigned to a region of the "Other" type and a sixth phase of association Set6 assigns a grey level having a digital value equal to 255 to the pixel I_r(x,y) of the grey-level image temp_Img_Reg; the association operation then terminates and a new pixel is taken into consideration.

0078As can be noted, the conditions for belonging to the regions can be briefly summarized as follows: <ul id="ul0005" list-style="bullet" compact="compact"><li>conditions whether or not the parameters of the pixel belong to a given interval, or</li><li>mathematical relationships between parameters of one and the same pixel, or</li><li>any combination of the aforementioned conditions and combinations.</li></ul>

0079In figure 5 the reference Img_RGB indicates an output image of the colour reconstruction block Col_Rec, while Img_s and Img_Reg indicate the corresponding output images of, respectively, the segmentation block Img_Seg and the block Reg_Detect that divides the image into regions. As may be noted, the image Img_Reg is divided into four regions, respectively, from top to bottom: of the "Sky" type (r5), the "Farthest mountain type (r3), the "Near mountain" type (r1) and the "Land" type (r4).

0080Coming back to the block diagram of Figure 1, a class decision block makes it possible to associate with the image Img_Reg a class I_Class forming part of a predefined set of image classes CL₁, CL₂, ..., CL_n. In a particularly preferred embodiment the set of predefined classes includes the following classes: CL₁ = "outdoor with geometric appearance", CL₂ = "outdoor without geometric appearance (or panoramas)", CL₃ = "indoor" and similar.

0081It has been found that each class may be characterized by means of an appropriate simplified model obtained on the basis of a statistical observation of some characteristics typically present in images belonging to the class.

0082In a particularly preferred embodiment a predefined Class CL₁ is characterized by a respective model comprising one or more predefined sequences (i.e. successions in accordance with a particular disposition) of regions that can be typically observed along a predetermined scanning direction in images belonging to this class. In particular, each predefined sequence includes a string of identifying labels (i.e. data) assigned to regions typically found along the predetermined scanning direction.

0083For example, considering the following table of labels identifying correspondences between regions and grey levels: <tables id="tabl0002" num="0002"><table frame="all"><tgroup cols="3" colsep="1" rowsep="1"><colspec colnum="1" colname="col1" colwidth="52.50mm" /><colspec colnum="2" colname="col2" colwidth="52.50mm" /><colspec colnum="3" colname="col3" colwidth="52.50mm" /><thead valign="top"><row><entry namest="col1" nameend="col1" align="center"><b>REGION</b></entry><entry namest="col2" nameend="col2" align="center"><b>LABEL</b></entry><entry namest="col3" nameend="col3" align="center"><b>GREY LEVEL</b></entry></row></thead><tbody valign="top"><row><entry namest="col1" nameend="col1" align="left">R5 = Sky</entry><entry namest="col2" nameend="col2" align="left">s</entry><entry namest="col3" nameend="col3" align="left">0</entry></row><row><entry namest="col1" nameend="col1" align="left">R3 = Farthest mountain</entry><entry namest="col2" nameend="col2" align="left">M</entry><entry namest="col3" nameend="col3" align="left">50</entry></row><row><entry namest="col1" nameend="col1" align="left">R2 = Distant mountain</entry><entry namest="col2" nameend="col2" align="left">M</entry><entry namest="col3" nameend="col3" align="left">80</entry></row><row><entry namest="col1" nameend="col1" align="left">R1 = Nearby mountain</entry><entry namest="col2" nameend="col2" align="left">M</entry><entry namest="col3" nameend="col3" align="left">100</entry></row><row><entry namest="col1" nameend="col1" align="left">R4 = Land</entry><entry namest="col2" nameend="col2" align="left">L</entry><entry namest="col3" nameend="col3" align="left">200</entry></row><row rowsep="1"><entry namest="col1" nameend="col1" align="left">R6 = Other</entry><entry namest="col2" nameend="col2" align="left">X</entry><entry namest="col3" nameend="col3" align="left">255</entry></row></tbody></tgroup></table></tables> it was found that the sequences of regions that could be observed with a certain statistical frequency in images belonging to the class CL₁ = "outdoor without geometric appearance (or panoramas)" are expressed by the following strings of identifying labels: <ul id="ul0006" list-style="bullet" compact="compact"><li>"s"</li><li>"sm"</li><li>"sl"</li><li>"sml"</li><li>"m"</li><li>"ml"</li><li>"l"</li><li>"sms"</li><li>"smsl"</li><li>"msl"</li><li>"mls"</li><li>"Is"</li></ul>

0084The label "s" placed after the labels "m" and/or "1" indicates situations in which a generic water surface is present in the lower part of the image.

0085It has also been found that the images belonging to the predefined class CL₂ = "outdoor with geometric appearance" present region sequences that do not belong to the set of sequences listed above, but nevertheless comprise a region of type s, i.e. "Sky". in the upper part.

0086Lastly, the remaining regions are characteristic of images belonging to the class CL₃ = "indoor".

0087Advantageously, in the class decision block, which is shown in greater detail in Figure 6, there is carried out a phase f₁ for acquiring information from the image Img_Reg divided into regions to obtain information about the regions present in it.

0088Preferably, the information acquired in this phase f₁ concerns: the number of transitions between regions and/or the type of regions present and/or the respective order of disposition in accordance with a predetermined scanning direction.

0089In an even more preferred embodiment, the information acquisition phase f₁ comprises an operation S_Scan of acquiring one or more sequences S_j of regions from the image Img_Reg divided into regions. This operation comprises, for example, one or more scannings of the image divided into regions along one or more predefined directions. In greater detail, during the scanning along a particular direction the identifying labels assigned to the regions encountered during the scanning are memorized in a string S_j of characters.

0090Preferably, the scanning of the image is a scanning of the vertical type, from top to bottom for example, of the pixels belonging to one or more columns of the image. For example, the columns are selected at regular intervals starting from the top left-hand corner of the image. Hereinafter the present description will make reference to the scanning along the columns of the image without any limitation.

0091Preferably, the operation of acquiring the number N of acquired regions, i.e. the number N of columns subjected to scanning, is comprised between 5 and 10. In fact, it has been found that even though the acquisition of a large number N of sequences S_j provides a larger quantity of information about the image that is to be classified, it does not always increase the probability of having a correct classification. In many cases, in fact, the experiments that envisaged the acquisition of a large number of sequences S_j provided the same results as experiments carried out with a smaller number of acquired sequences.

0092Preferably, the grey levels associated with the pixels and the identifying labels encountered in the scanning direction are memorized during the scanning. For example, an identifying label is memorized in a string S_j whenever it differs from the last identifying label memorized in the string S_j. Obviously, the label associated with the first region encountered is always memorized. Even more preferably, a label is memorized in the string S_j whenever the number of pixels encountered during the scanning of a column exceeds a predetermined threshold value H_min that corresponds to a minimum number of pixels chosen in manner proportional to the height H of the image. Preferably, the threshold value H_min is substantially comprised between 5% and 15% of the vertical dimension H of the image. In a preferred embodiment, the information acquisition phase also comprises a counting operation J_Count along one or more predetermined directions of the number C_j of transitions, i.e. passages (or jumps), between regions of different types of the image divided into regions.

0093For example, the counting phase is carried out directly during the scanning of a sequence S_j of the image divided into regions by increasing a counter C_j every time a transition between different regions is detected. In this case a corresponding counter C_j can be associated with every string S_j, thus obtaining a data pair (S_j, C_j) for every scanning of the image.

0094For example, a transition is detected during the scanning along a column of the image divided into regions whenever there is read a value (grey level) of a pixel different from the last previously read value.

0095It should be noted that, given a highly fragmented image divided into regions, even though the number of labels in the acquired strings S_j may be limited on account of the control of the minimum number of pixels (H_min) , the value of the corresponding transition counter C_j may be high.

0096Figure 7 shows the image divided into regions of Figure 5 together with five columns c1, c2, c3 c4, c5 selected for scanning.

0097On the assumption that the columns c1, c2, c3 c4, c5 are scanned from top to bottom, the label strings Sj read and the number and the transitions counted by the corresponding counters C are shown column by column in the following table: <tables id="tabl0003" num="0003"><table frame="all"><tgroup cols="3" colsep="1" rowsep="1"><colspec colnum="1" colname="col1" colwidth="52.50mm" /><colspec colnum="2" colname="col2" colwidth="52.50mm" /><colspec colnum="3" colname="col3" colwidth="52.50mm" /><thead valign="top"><row><entry namest="col1" nameend="col1" align="center"><b><u>Column</u></b></entry><entry namest="col2" nameend="col2" align="center"><b>String</b></entry><entry namest="col3" nameend="col3" align="center"><b>Number of transitions</b></entry></row></thead><tbody valign="top"><row><entry namest="col1" nameend="col1" align="left"><b><u>Column c1</u></b></entry><entry namest="col2" nameend="col2" align="left">S₁="sml"</entry><entry namest="col3" nameend="col3" align="left">C₁=3</entry></row><row><entry namest="col1" nameend="col1" align="left"><b>Column c2</b></entry><entry namest="col2" nameend="col2" align="left">S₂="sml"</entry><entry namest="col3" nameend="col3" align="left">C₂=3</entry></row><row><entry namest="col1" nameend="col1" align="left"><b>Column c3</b></entry><entry namest="col2" nameend="col2" align="left">S₃="sml"</entry><entry namest="col3" nameend="col3" align="left">C₃=3</entry></row><row><entry namest="col1" nameend="col1" align="left"><b>Column c4</b></entry><entry namest="col2" nameend="col2" align="left">S₄="ml"</entry><entry namest="col3" nameend="col3" align="left">C₄=3</entry></row><row rowsep="1"><entry namest="col1" nameend="col1" align="left"><b>Column c5</b></entry><entry namest="col2" nameend="col2" align="left">S₅="ml"</entry><entry namest="col3" nameend="col3" align="left">C₅=2</entry></row></tbody></tgroup></table></tables>

0098Coming back to the block diagram of Figure 6, in the class decision block there is also carried out an analysis phase f₂ of the information acquired in the acquisition phase f₁. The association of one of the predefined classes CL₁, CL₂, CL₃ with the image divided into regions is obtained on the basis of the result of this analysis phase f₂.

0099In particular, the analysis phase f₂ includes a phase Cmp of comparing the information (S_j, C_j) obtained in the acquisition phase1 with the models that characterize the predefined classes.

0100The preferred embodiment in which information about the region sequences S_j and the transition counters C_j is acquired during the information detection phase f₁ also includes an operation S_Discard for selecting a subset of the N acquired sequences S_j comprising sequences S_j associated with a respective transition counter C_j containing a value smaller than or, at the most, equal to a predetermined limit value C_th. Furthermore, the non-selected sequences are preferably discarded. For example, it has been found in the course of experimental tests that good results are obtained with a limit value C_th approximately equal to ten.

0101The comparison operation Cmp is such as to verify whether the selected acquired sequences, i.e. the non-discarded sequences, are present among the predefined sequences of a model characterizing a predefined class.

0102Preferably, a predefined class CL₁, CL₂, CL₃ is associated with the image whenever among the predefined sequences of the model that characterizes the predefined class there is present a number of selected sequences greater than a predetermined percentage of the total number N of acquired sequences S_j.

0103In the example of Figure 6 the comparison operation Cmp includes a first control operation Cmp1 in which there is determined the number N of the acquired and not discarded sequence S_j that are comprised in the set of predefined sequences belonging to the model of the Class CL₁ = "outdoor without geometric appearance (or panoramas)". Whenever the determined number N₁ is greater than or equal to a predetermined percentage of the total number N of acquired sequences S_j, a first association operation Set CL₁ assigns the predefined class CL₁ to the image. In pseudo code, whenever:<maths id="math0003" num=""><math display="block"><mrow><msub><mrow><mtext>N</mtext></mrow><mrow><mtext>1</mtext></mrow></msub><msub><mrow><mtext> ≥ R</mtext></mrow><mrow><mtext>1</mtext></mrow></msub><mtext> x N</mtext></mrow></math><img file="EP1507232A2_D0003.tif" /></maths>    then I_Class=CL₁    wherein R₁ is a number comprised between 0 and 1. The value of R₁ plays a fundamental part in associating the predefined class CL₁, because it furnishes a measure of the precision level of the entire method. It is obvious to note that low values of R₁ (smaller than 0.5, for example) make it possible to classify an image as belonging to the predefined class CL₁ with greater "ease". In these cases one therefore risks "taking for good" even images that in reality are not so.

0104On the other hand, utilizing a high value of R1 (greater than 0.8, for example) possibly risks an excessive thinning of the tolerance margin that has to be assured for a heuristic method.

0105From the acquired experience and the analyzed results it was found that values of R₁ substantially comprised between 0.6 and 0.8 make it possible to obtain sufficiently accurate and trustworthy results.

0106Whenever the number N₁ does not satisfy the above condition, a second control operation Cmp2 is carried out to determine the Number N₂ of the acquired and non-discarded sequences S_j that are comprised in the set of predefined sequences belonging to the model of the class CL₂ = "outdoor with geometric appearance". When the determined number N₂ is greater than or equal to a predetermined percentage of the total number N acquired sequences S_j, a second association operation assigns the predefined Class CL₂ to the image. In pseudo code, when:<maths id="math0004" num=""><math display="block"><mrow><msub><mrow><mtext>N</mtext></mrow><mrow><mtext>2</mtext></mrow></msub><msub><mrow><mtext> ≥ R</mtext></mrow><mrow><mtext>2</mtext></mrow></msub><mtext> x N</mtext></mrow></math><img file="EP1507232A2_D0004.tif" /></maths>    then I_Class = CL₂    wherein R₂ is a number comprised between 0 and 1.

0107Since in this case the predefined class includes sequences of regions that begin with the region "Sky", the tolerance margin determined by the parameter R₂ is very ample, and experimental results have shown that values of R2 substantially comprised between 0.1 and 0.5 make it possible to obtain sufficiently accurate and trustworthy results. In a particularly preferred embodiment, R2 is substantially equal to 02.

0108When the number N₂ does not satisfy the above condition, a third association operation Set_CL3 is carried out to assign to the image the predefined class Cl₃ = "indoor", which therefore contains all the sequences of regions not comprised in the models that characterize the predefined classes CL₁ and CL₂.

0109The proposed method has such advantages as automaticity, computational simplicity and efficiency.

0110Automaticity is a very important property of any image processing procedure. Indeed, the construction of a technique capable of providing reliable results by eliminating interaction with the human user represents a step forward in the global process of automation of the presently available technologies. Furthermore, it implies the possibility of aggregating image classification with other automatic processes for the purpose of creating advanced multi-process systems.

0111Simplicity and efficiency are two characteristics that do not always go hand in hand. In the method in question the problem of image classification is resolved by having recourse to a multi-stage system made up of a series of functional blocks that are relatively simple because they are implemented by means of algorithms based on rather elementary heuristic characterizations and statistical properties. Efficiency depends primarily on having used input images in CFA format and having made provision for colour reconstruction with sub-sampling, thereby making it possible to work with images having dimensions equal to half the original ones, and having used the least number of possible operations (see, for example, the simplicity of the class decision block).

0112The method in accordance with the present invention may be directly utilized, for example, in such portable or fixed image acquisition devices as: digital photo/video cameras and the like, multimedia communication terminals capable of acquiring and/or receiving digital images. Furthermore, the method in accordance with the present invention can be executed as a program or a program portion in a processing device, a personal computer for example. This program can be memorized in any appropriate manner, for example in a memory or on such a memorization support as a floppy disk or a CD ROM.

0113Obviously, with a view to satisfying contingent and specific needs, a person skilled in the art will be able to introduce numerous modifications and variants into the classification method described above, though all contained within the protection limits as defined by the claims attached hereto.