Object detection in an image
The Riesz transform-based method addresses the challenge of detecting low resolution and camouflaged objects by utilizing structure tensors and eigenvalues to enhance object detection accuracy.
Patent Information
- Application Number
- FR2022011917
- Authority / Receiving Office
- FR · FR
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2022-11-16
- Publication Date
- 2025-01-03
- Estimated Expiration
- 2042-11-16
AI Technical Summary
Current object detection algorithms struggle to identify low resolution objects or those with colors similar to the background in images.
A method involving the application of a Riesz transform to image channels to calculate structure tensors, determine eigenvalues, and process these values through an object detection neural network, enhancing detection capabilities.
The method effectively detects low resolution and camouflaged objects by leveraging Riesz transform-based structure tensors and eigenvalues, improving object detection accuracy.
Smart Images
Figure 00000019_0000 
Figure 00000020_0000 
Figure 00000021_0000
Abstract
Description
Title of the invention: Object detection in an image FIELD OF INVENTION
[0001] The present invention relates to the field of image processing, more specifically to the detection of objects in an image and, more particularly, to the detection of low-resolution objects or camouflaged objects in an image. STATE OF THE ART
[0002] To detect objects in an image, different algorithms can be used, including neural networks such as, for example, YOLO, R-CNN, Fast R-CNN or Faster R-CNN.
[0003] Current techniques for detecting objects in an image do not allow the detection of low-resolution objects or objects with colors similar to the background of the image. Description of the invention
[0004] One object of the invention is to improve the detection of objects in an image. More specifically, the invention aims to enable the detection in an image of objects with low resolution or having colors similar to the background of an image.
[0005] According to a first aspect, a computer-implemented method for detecting objects in an image comprising one or more channels is proposed, comprising the following steps: a) for at least one channel among the image channel(s), calculation of two Riesz components by applying a Riesz transform to the channel and determination, from the two Riesz components, of a structure tensor having the Riesz transform as its operator; b) determination of an image characterization matrix by summation of the structure tensor(s) determined in step a); c) determination of at least one characteristic of the image from at least one eigenvalue of the image characterization matrix; d) processing of the feature by an object detection neural network.
[0006] According to advantageous and non-limiting features, taken alone or in any combination: - The structure tensor determined in step a) for an image channel is expressed as follows: ic\,2 (R^ BEEP g*1 cxc _v in which jiÿ 7g1.2 R^GC X1VC y Mï V 'I C y) (V y ) corresponds to the structure tensor along a direction 0 for a pixel Ic(x, j) of the channel C of the image, and correspond to the terms of the structure tensor c 1.1' c 1.2 c 2.2 T^Uy)- and correspond to the Riesz components along x and axes V C x X y y respectively resulting from the Riesz transform applied to channel C of the image along a direction ¢, C corresponds to one of the channels Cl to CN in the image, N is an integer greater than or equal to 1, 0 corresponds to the calculation direction of the structure tensor and is understood to between -ji and ir and G corresponds to the kernel order of the Riesz transform;
[0007] - the determination of at least one characteristic of the image includes the determination of the eigenvalues of the image characterization matrix and selection of the largest of said eigenvalues;
[0008] - in step c), a feature of the image is determined by calculating: A ~ ^^T^Tdans lecluel' ^2 correspond to the eigenvalues of the image characterization matrix;
[0009] - the method further comprises determining at least one local orientation of the image from an eigenvalue of the structure tensor determined in step a) for an image channel and step d) includes the processing of local orientation by the object detection neural network;
[0010] - the process includes, before step b), at least one repetition of step a) in modifying a direction of the Riesz transform, the structure tensors determined at each of the iterations of step a) for the same image channel being added together in a step al);
[0011] - at step a), the Riesz transform is a Gaussian kernel Riesz transform.
[0012] - the image is a color image comprising three channels, including a red channel, a a blue channel and a green channel.
[0013] According to a second aspect, an object detection system in an image is proposed, comprising a processor configured to implement the detection method presented previously.
[0014] According to a third aspect, a computer program product is proposed comprising instructions which, when the program is executed by a computer, lead the latter to implement the detection method presented previously.
[0015] According to a fourth aspect, a computer-readable recording medium is proposed comprising instructions which, when executed by a computer, lead the computer to implement the detection method described above. DESCRIPTION OF THE FIGURES
[0016] Other features and advantages of the present invention will become apparent from the following description of a preferred embodiment. This description will be given with reference to the accompanying figures, including:
[0017] - [Fig.1] illustrates an object detection system in an image;
[0018] - [Fig.2] represents the steps of an object detection process in an image;
[0019] - the [Fig.3] is an image acquired by a camera;
[0020] - [Fig.4] represents the phase of the image of [Fig.3];
[0021] - [Fig.5] represents the amplitude of the image of [Fig.3];
[0022] - the [Fig.6] is a second image acquired by a camera;
[0023] - Figures [Fig. 7], [Fig. 8], and [Fig. 9] represent local orientations of the image of the [Fig.6];
[0024] - Figure 10 schematically illustrates the data flow within the process of object detection in an image. DETAILED DESCRIPTION OF THE INVENTION System
[0025] A system 1 for detecting objects in an image is illustrated in [Fig. 1]. The system 1 comprises a computer 2 configured to implement the method for detecting objects in an image which will be presented subsequently.
[0026] Computer 2 is, for example, a processor.
[0027] System 1 is configured to process one or more images. System 1 is therefore configured to receive images.
[0028] System 1 can be configured to receive images from external sources such as an image acquisition device like a camera or a data storage device such as a computer memory.
[0029] According to a certain embodiment, the system 1 includes image acquisition means 3 and the system 1 is configured to process the images acquired by the image acquisition means 3.
[0030] The images can, for example, be photos or videos.
[0031] The images can be processed by system 1 in real time. By "real time," we mean that the images are processed by system 1 directly after their acquisition, that is, preferably less than 1 ms after their acquisition and, even more preferably, less than 0.42 ms after their acquisition. In other words, system 1 takes the images as input directly after their acquisition. In this case, system 1 is configured to receive images from an image acquisition device such as a camera.
[0032] According to another embodiment, the images are not processed in real time and are acquired some time before being processed by system 1 (for example, several minutes or several hours). In other words, the images are provided as input to system 1 some time (for example, several minutes or several hours) after their acquisition. In this case, system 1 is configured to receive images from a data storage device such as computer memory.
[0033] System 1 can be configured to receive both real-time images and images that are not real-time.
[0034] The images may, for example, represent landscapes, buildings, people, animals, vehicles, etc...
[0035] The images have one or more C channels. In other words, each image comprises a plurality of pixels having one or more C channels.
[0036] The image can be a color image with at least three channels: a red channel, a green channel and a blue channel.
[0037] The image may be a greyscale image with a single greyscale channel.
[0038] The image may present other types of channels corresponding to various colors or various spectra, for example infrared.
[0039] The image can be a two-dimensional image or a multi-dimensional image. The image has dimension N. Typically, the image is a two-dimensional image acquired by two-dimensional image acquisition means. This is the case of a two-dimensional image that will be primarily processed subsequently. Method
[0040] With reference to [Fig.2], a method for detecting objects in an image is proposed.
[0041] The method is implemented by computer, typically by the object detection system 1 presented previously.
[0042] The method preferably includes a preliminary step a0) of obtaining an image. The image can be directly acquired by image acquisition means 3 of system 1. The image can also be provided by external sources.
[0043] The image comprises one or more pixels. The method that will be described aims to be applied to a plurality of pixels of an image, or even to each pixel of the image, so as to be able to detect objects in the image.
[0044] The process then includes a step a) of obtaining at least two Riesz components by applying a Riesz transform to the image.
[0045] The Riesz transform corresponds to the generalization of the Hilbert transform.
[0046] Preferably, the Riesz transform used is a kernel Riesz transform Gaussian.
[0047] The Riesz transform is calculated for at least one channel C of the image pixel.
[0048] The Riesz transform is calculated for an order G with G an integer greater than or equal to 1.
[0049] The Riesz transform is calculated at least along one direction with 0 between 0 and jt.
[0050] The Riesz transform is calculated for an image of dimensions N with N an integer greater than or equal to 2.
[0051] Thus, the Riesz transform of order G of a pixel I(x, y) of a two-dimensional image I along a direction 0 of the transform and a channel C of the image can be written as follows:
[0052] pr[ / (x, j*=k^°(X, y) =
[0053] With I(x, y) corresponding to the pixel of the image with coordinate x along an x-axis and coordinate y along a y-axis,
[0054] C corresponding to one of the channels Cl to CN of pixel I(x, y),
[0055] Rc6 x (x' y )= ü (pG t1 (y ) 1cos^ ) et ^cGyXx,y) corresponding to the components of Riesz along the x and y axes respectively of the Riesz transform of order G along the 0 direction applied to a channel C of the pixel I(x, y) ,
[0056] Jq j cl j corresponding to gradients along the x and y axes, respectively,
[0057] p6]. . ] which corresponds to the Riesz operator of order G,
[0058] G corresponding to the order of the kernel of the Riesz transform and being an integer greater than or equal to 1,
[0059] fi corresponding to the direction of the Riesz transform and being between - ir and ji,
[0060]
[0061]
[0062]
[0063]
[0064]
[0065]
[0066]
[0067]
[0068]
[0069]
[0070]
[0071]
[0072]
[0073]
[0074]
[0075]
[0076]
[0077]
[0078]
[0079]
[0080] In the time domain, for a channel C, the x-axis component of the Riesz transform calculated with a kernel of order G along a direction A, denoted xj) 'Can be s ^cr^re as follows: J) =Jp!^ / c(xr)dr with Ic(xt) corresponding to channel C of the translated image of t = [rj, t2 ] , '1 corresponding to the sampling step along the x-axis, r~ corresponding to the sampling step along the y-axis, dr = dr id~-, and P2 corresponds to the set of pairs of real numbers. In the frequency domain, for a channel C, the component along the x-axis of the Riesz transform calculated with a kernel of order G along a direction ¢, denoted ( X j) ' Can be written as follows: A $£> RC k(kï>k2) with k { which corresponds to a coordinate (i.e., a frequency) along an axis in the frequency domain and k2 (i.e. a frequency) which corresponds to a coordinate along an axis in the frequency domain. This component can be calculated using the following formula: RC k ( kb k2) - J( G h'(kbk2) with j corresponding to the complex number, expressed as follows: eXp( — with * corresponding to the imaginary unit, jk} ) 9U' corresponds to the Fourier transform of the channel C of the pixel / (x, y) in the frequency domain. Respectively, in the time domain, for a channel C the component along the y-axis of the Riesz transform calculated with a kernel of order G along a direction ¢, denoted j^^G yy, can be written as follows: y} = with Ic ( X - T ) corresponding to channel C of the image translated from t - [rq, t2 ], corresponding to the sampling step along the x-axis, T- corresponding to the sampling interval along the y-axis, dT = dTidr2 and p2 corresponding to the set of pairs of real numbers.
[0081]
[0082]
[0083]
[0084]
[0085]
[0086]
[0087]
[0088]
[0089]
[0090]
[0091]
[0092] In the frequency domain, for a channel C, the component along the y-axis of the Riesz transform calculated with a kernel of order G along a direction A, denoted (xj) ' Can be written as follows: This component can be calculated using the following formula: SI I \ T SI 1 Rc k (^1^2) G ^2) 2 (AT+kri Step a) of the process further includes determining, from the Riesz components ^c'y(xy) >, a structure tensor having operator the Riesz transform of order G denoted pG[. . ]. For the sake of brevity, R^À^y^'R^',(x,y) will be denoted by and , respectively. It is therefore understood that grfp corresponds to C x 5' C x Typically, a structure tensor is a matrix derived from the gradient of a function. In the case of this invention, the structure tensor is a matrix derived from the Riesz transform. As will be explained later, the structure tensor with the Riesz transform as its operator allows for the calculation of image features that are very useful for object detection. It has been observed that the structure tensor with the Riesz transform as its operator characterizes image textures more effectively than a classical structure tensor, i.e., one with the gradient as its operator. As explained, the structure tensor has the Riesz transform as its operator. More precisely, the structure tensor has the Riesz transform of order G as its operator, where G is an integer greater than or equal to 1. The structure tensor is preferably determined for each pixel of the image. The structure tensor is determined for a direction ¢. More precisely, the direction corresponds to the direction for which the Riesz transform of the pixel is calculated. Indeed, the unoriented Riesz transform of a pixel can be "rotated" to obtain the oriented Riesz transform of the pixel along a direction ¢. By "unoriented", we mean that the kernel used for the calculation of the transform is a zero-orientation Gaussian filter. The Riesz transform for an unoriented N-dimensional image of order G of the channel C of a pixel / (x, y) can be written as follows:
[0093] IP^lKx.yïjJ PG[r(x,y)]c= P^lHx.y)^ y) ]C|
[0094] with n a dimension between 0 and N -1.
[0095] In the case of a two-dimensional image, N is equal to 2 and we have ^[ / (^)1,=1^)^¾^))^4, y)-
[0096] In the frequency domain, p("," j) ] is denoted p(n. g)j jAr ) ] with '
[0097] ^[7(^^)] = lvp lj j£ 1
[0098] Multiplying ( ) Nn by corresponds to the derivative partial of • Dividing by the norm of allows us to retain that the information concerning the W phase of the Riesz transform of the image.
[0099] We can say that pG is an "allpass" type filter bank with, for each filter, a directional kernel p"G).
[0100] To return to the matter of rotating the Riesz transform, as explained, the Riesz transform is orientable, i.e., the kernel used can be rotated. In other words, the Riesz transform of a pixel can be oriented towards a direction ¢ and this oriented Riesz transform can be defined by a linear combination of all the components of the unoriented Riesz transform with the values of the rotation matrix.
[0101] y)]^ = Aÿ€ PG[Z(x,j)]c ' 4^ A(),0 4^ A(U a0,GP^fZCx, y) ]c PG[Z(x,y)]^ = P(”-G)[Z(x,y)]JA An# a»,g ^[Z(XJ)]CA# \ Av-io A^ A / ^[ZÇxyHj pW,.< a / ' pfet
[0102] with $ between 0 and G.
[0103] Thus, by multiplying the unoriented Riesz transform of a pixel by a rotation matrix, we obtain the oriented Riesz transform of the pixel along the direction ¢.
[0104]
[0105]
[0106]
[0107]
[0108]
[0109]
[0110] [YES]
[0112]
[0113]
[0114]
[0115]
[0116]
[0117]
[0118]
[0119] The definition of the rotation matrix is derived from the frequency expression of and thus the following equations: An,G) r-^ (J(cA0)A™n^ * — y uKN-n) ! 11^1° -jkA^w -ik,r Ç»(V) )”'!( - jkAYùnl.t) Y'A -jk,^'"' I / A'G v" V'V» z . m.[n{N -7^2) 1 "0 / / / , ] (^rrs'(^)) / • / < \ \ n-ntf+m. _yH y .X „ i NG (-jk^ ~ ^mf=0^m2=0 \l(m}+m2^(G-mrm2J _____________—----------- ni, „t (N-nf. . . ,, . N^m+rn-, , . , , x --------(-1) -(sln(¢)) with w corresponding to (kp k2 ), being between 0 and n and m2 being between 0 and N - n. A (n, Nn) R ^I^a^SV-N+n) 1 1 y \IOA <P) ) f SIU [ ) ) with 1 being between 0 and 1 being between max(Q,lG + g) and min(l, g). Finally, for a pixel Ic(x, y) of a channel C, the Riesz components oriented along the direction are expressed, for the x and y axes respectively, ü (p g R(*, j) 1^)^ et
[0120]
[0121]
[0122] The structure tensor can be written as follows: A'1S 7^,.2] 1«*^ J with Çx y) corresponding to the structure tensor along a direction for a pixel icA- y) of channel C of the image I(.x, y) and 71^ and 6 1.1' c 1.2 c 2.2 corresponding to the terms of the structure tensor for a pixel Ic(x> J) of the channel C of the image.
[0123] It should also be noted that the structure tensor can be expressed as a spatial correlation matrix of the Riesz components along the x and y axes, respectively. In this case, the structure tensor can be written as follows:
[0124] r^cj R*'*?5! T^^x^y)- J'1 , ]== ' = c [3* I2 T(: ..] ] £,(^( / (^ / ))01^^(^ / )}^^
[0125] with W corresponding to a local window of size wi x w2 around the pixel channel
[0126] Here, the eigenvalues of this structure tensor in the form of a correlation matrix characterize the correlation of the channel C of the pixel (x, y) with its neighborhood, i.e. with the pixels neighboring the pixel (x, y) in the window W.
[0127] This simplified structure tensor, even though it is calculated for a given direction using the first-order Riesz transform p1 as the operator, allows, for a channel C of the image, for a good characterization of the image pixels compared to a structure tensor using the gradient as the operator. Indeed, the characteristics of the image pixels that can be obtained using the tensor are particularly useful in the context of object detection using a neural network, as will be explained later.
[0128] Step a) is implemented for at least one channel of the image pixel. Preferably, step a) is implemented for a plurality of image pixel channels, or even for all image pixel channels. Consequently, step a) yields one or more structure tensors, each tensor being associated with a channel.
[0129] Thus, the Riesz transform can be said to be multispectral when applied to several channels of an image. Furthermore, the Riesz transform can be said to be monogenic since it produces a monogenic signal, which is inherent to the Riesz transform.
[0130] Preferably, step a) can be repeated several times for the same channel by modifying one direction of the Riesz transform and the structure tensors determined at each iteration of step a) for the same channel of the image pixel are summed together. In other words, for the same channel of the image pixel, step a) of the process can be repeated for a plurality of directions. Then, at the end of all the iterations of step a) for the same channel, in step a1), the structure tensors obtained for the different directions (but for the same channel) are summed together according to the following formula: Te(Xy) - (xy) • The directions chosen for the calculation of the Structure tensors are chosen such that all image textures are characterized. In practice, we choose values of 0 between [-TT, 71] with a step size of, preferably a step size of 75.
[0131] It is therefore understood that, if one wishes, for example, to calculate the structure tensor of a channel along two directions, two iterations of step a) for the same channel are carried out. Thus, for a channel C, a first iteration of step a) results in a structure tensor along a first direction for channel C. A second iteration of step a) results in a structure tensor along a second direction for channel C. At the end of these two iterations, the two tensors are added together to obtain a single structure tensor of channel C of the image pixel, which therefore combines the information from the Riesz transform of the image pixel along two different directions.
[0132] Two embodiments can thus be distinguished. According to a first embodiment, structure tensors are calculated along the same direction for one or more different channels. Thus, step a) is repeated the same number of times as the number of channels for which a structure tensor is to be calculated. These different iterations make it possible to obtain as many structure tensors as there are channels considered.
[0133] According to a second embodiment, structure tensors are calculated along a plurality of different directions and for one or more different channels. Several steps a) are implemented for the same channel but for different directions, resulting in a plurality of structure tensors, each calculated for the same channel but for different directions. Then, in one or more steps a1), the structure tensors for the same channel (but along different directions) are summed to obtain one structure tensor per channel considered.
[0134] In all cases, it is understood that, at the end of steps a) (according to the first embodiment) or al) (according to the second embodiment), one arrives at one or more structure tensors, each structure tensor being calculated for a channel different from the other structure tensors.
[0135] The method then includes a step b) of determining an image characterization matrix by summing the one or more structure tensors determined in step a) for one or more channels of the image pixel.
[0136] The image characterization matrix therefore preferably groups information about the image calculated for a plurality of channels of the image pixel. The image characterization matrix corresponds to a sum of structure tensors, each structure tensor being the structure tensor for a certain channel C.
[0137] The image characterization matrix can be written as follows:
[0138] Mi 2 M22 'Ce[Cl..CN} Tc(.x,y)
[0139] with M(x, y) the image characterization matrix for a pixel I(x, y) of the image,
[0140] Myj, M\2 and ^2.2 the terms of the image characterization matrix M(x, y),
[0141] C corresponding to one of the channels Cl to CN of the pixel of the image Z(x, y) with N corresponding to an integer greater than or equal to 1.
[0142] Preferably, the image characterization matrix corresponds to the sum of the structure tensors for each channel of the image pixel. For example, for a pixel having three channels, including a red channel, a green channel, and a blue channel, a structure tensor is calculated for each of the three pixel channels, and the image characterization matrix corresponds to the sum of the structure tensors for each of the three channels.
[0143] The image characterization matrix can correspond to the sum of the structure tensors for a selection of the image pixel channels.
[0144] Furthermore, it should be noted that each tensor is calculated along a given direction. Thus, for the same pixel channel, several structure tensors can be calculated, each being calculated along a given direction.
[0145] In the case where it is desired to calculate the structure tensor of a pixel channel for a plurality of directions ¢, the image characterization matrix can be written as follows: ^1.2 ^2,2 “ ^Ce^C L. ,CV}^-Væ{0. . y )
[0146] M(x,y) =
[0147] Preferably, as explained above, each term (xj) was calculated in accordance with step al).
[0148] Then in a step c), at least one feature of the image pixel is determined from at least one eigenvalue of the image characterization matrix. Preferably, at least one feature of at least one image pixel is determined. Preferably, at least one feature is calculated for all the image pixels.
[0149] Preferably, a determined pixel characteristic of the image is a phase 7' of the Riesz transform of the image. This phase V7 corresponds to the largest eigenvalue of the image characterization matrix.
[0150] The eigenvalues of the image characterization matrix tensor are denoted by and 22. The eigenvalue A can be written as follows: 5 ( M 11"*” 4 / 2 2 + ( 4 / 1 1 -4^2,2 ) ~ + 4A / j 2^ ) '
[0152] The eigenvalue 42 can be written as follows:
[0153] h = I ( 4 / U + 4¾ - + ) •
[0154] The phase V' of one or more pixels in an image allows for the identification of objects within that image. In particular, the phase V highlights the contours of the image. For example, starting from an image illustrated in Figure 3 (acquired from a camera, for instance), it is possible to obtain a phase V' image illustrated in Figure 4, which is generated from phases W calculated for the pixels of the image. Several objects, such as people P1, P2, and P3, can be identified in this phase V' image. In the image in [Fig. 3], people P1, P2, and P3 can be considered low-resolution objects.
[0155] Preferably, a feature calculated in step c) corresponds to an amplitude A of the Riesz transform of the image pixel. The amplitude can be calculated as follows: [° 156 1 A =^71
[0157] with, as a reminder, and 42 the eigenvalues of the image characterization matrix.
[0158] [Fig.5] illustrates the pixel amplitude of the image of [Fig.3] and also highlights persons PI, P2 and P3.
[0159] Amplitude and phase are complementary information in the context of object detection. Thus, preferably, step c) of the method includes determining the amplitude and phase of the image pixel.
[0160] Advantageously, other characteristics of the image pixel are determined. More precisely, preferably, the method includes a step c1) in which at least one local orientation of the image pixel for a certain channel C is determined from an eigenvalue of the structure tensor determined in step a) for said certain channel. A local orientation is denoted c.
[0161] To this end, step cl) includes determining, from the Riesz components for channel C, a channel characterization matrix for channel C having the Riesz transform as its operator, said channel characterization matrix for channel C being determined by calculating:
[0162] Mc(x, y)
[0163] with Mc(x, y) the channel characterization matrix for a given channel C and
[0164] Mc p 2 and Mc^ the terms of the channel characterization matrix.
[0165] In other words, to calculate a channel characterization matrix, for each channel C one or more structure tensors are calculated, each structure tensor being calculated along a direction 0. It is understood that the number of tensors summed to calculate the channel characterization matrix corresponds to the number of directions 0 along which it is desired that the tensors be calculated.
[0166] The local orientation corresponds to the largest eigenvalue of the structure tensor for the channel C. The eigenvalues and of a channel characterization matrix Mc{x, y ) are calculated similarly to the eigenvalues and ^2 of the characterization matrix M ( x, y ).
[0167] The eigenvalue 41C can be written as follows:
[0168] ^ic - I (.1+^2,2 + ).
[0169] The eigenvalue ^2C can be written as follows:
[0170] = 4 ( Mr + Mr - \l ( 12 + ■ zy ^11^22 V * 1 1 ^2 c 12 /
[0171] Local orientations provide additional information about the image pixel. In particular, local orientations have the advantage of facilitating the detection of a camouflaged object, i.e., the color of the object is similar to the color of the image background.
[0172] The image illustrated in Figure 6 (acquired from a camera, for example) represents a person P4 who blends into the background, i.e., it is difficult to distinguish person P4 from the background. Figures 7 to 9 illustrate, respectively, the images with local orientations 6r, 0v, and calculated from the local orientations of the image pixels, which correspond respectively to the red, green, and blue channels of the image.
[0173] We note that person P4 is identifiable in these images.
[0174] Preferably, for each channel C of the image, a local orientation is calculated. Thus, for example, for an image including the three channels red, green and blue, the local orientations 9r, 0v and 0B are calculated as in the example illustrated in figures 7 to 9.
[0175] To summarize, the equation below summarizes how to obtain image characteristics as previously presented: Riesz P°ÏI(*, J)]*, pWj)]*, PWj) ]*! Structure tensors and characterization matrix(ies) HAS'. V' Gold Og
[0177] The method then includes a step d) of processing at least one feature by an object detection neural network. The neural network can be any type of neural network suitable for detecting objects in an image. For example, the neural network can be a YOLOV, R-CNN, Fast R-CNN, or Faster R-CNN.
[0178] In other words, step d) consists of providing the neural network with at least one feature of the pixel from the previously computed image as input. More precisely, preferably, for an image, at least one feature is provided as neural input for each pixel of the image.
[0179] The neural network was learned from a feature base extracted from images preferably representing low-resolution and / or camouflaged objects.
[0180] The neural network is configured to detect objects in the image based on this feature. More precisely, the neural network architecture is advantageously adapted to process the feature and consequently detect objects in the image.
[0181] Preferably, the amplitude A and phase W of the image are provided as input to the neural network and processed by the neural network. These two characteristics allow for very good object detection in an image, particularly of low-resolution objects and / or camouflaged objects. More specifically, advantageously, the amplitude A and phase V' of each pixel of the image are provided as input to the neural network.
[0182] Thus, the neural network preferably comprises at least two input branches, each branch comprising a plurality of processing blocks to process one or the other of the amplitude and phase of the image. Then the output data from each of the branches are concatenated to give a single output data which is then processed in other processing blocks of the network.
[0183] According to a certain embodiment, one or more local orientations are also provided as input to the neural network for object detection. Providing the neural network with amplitude, phase, and one or more local orientations as input allows, in particular, the detection of objects that are difficult to detect, such as low-resolution and camouflaged objects.
[0184] At the end of the process, objects are detected in the image. By "detected", we mean that at least the type of object and the location of the object are detected.
[0185] To summarize, as illustrated in [Fig. 10], according to one embodiment, the method of detecting object(s) in an image first includes the determination of at least one feature of the image by system 1 and then the provision of the feature as input to a neural network RN for object(s) detection to result in the detection of one or more objects in the image. Product, program, and recording medium
[0186] A computer program product is also proposed which includes instructions which, when the program is executed by a computer, cause the computer to implement the object detection method as previously described.
[0187] In addition, a computer-readable recording medium is proposed comprising instructions which, when executed by a computer, cause the computer to implement the object detection method as previously described.
[0188] The invention is not limited to the embodiment described and shown in the accompanying figures. Modifications remain possible, particularly with regard to the constitution of the various technical features or by substitution of technical equivalents, without departing from the general principles.
Claims
Claims
1. A computer-implemented method for detecting objects in an image comprising one or more channels, comprising the following steps: a) for at least one channel among the channel(s) of the image, calculating two Riesz components by applying a Riesz transform to the channel and determining, from the two Riesz components, a structure tensor having the Riesz transform as operator; b) determining an image characterization matrix by summing the structure tensor(s) determined in step a); c) determining at least one characteristic of the image from at least one eigenvalue of the image characterization matrix;d) processing the characteristic by an object detection neural network, the method comprising before step b) at least one reiteration of step a) by modifying a direction of the Riesz transform, the structure tensors determined at each of the iterations of step a) for the same channel of the image being added together in a step a1).;
2. Method according to claim 1, in which the structure tensor determined in step a) for a channel of the image is expressed according to: frf» rf [ C'f 7^.,) = . iJ / 2 = J Tc 1,2 T^\ R*? R?? j which ( xy ) corresponds to the structure tensor along a direction 0 for a pixel Ic(x, y) of the channel C of the image, t^g ^.G cj correspond to the terms of the tensor of c 1,1' c 1,2 c 2.2 structure T^(x,y)' and j^p correspond to the Riesz components along cxc J of the axes x and y respectively resulting from the Riesz transform applied to the channel C of the image along a direction C corresponds to one of the channels Cl to CN of the image, V is an integer greater than or equal to 1, 0 corresponds to the direction of calculation of the structure tensor and being between - ir and ir and G corresponds to the order of the kernel of the Riesz transform.
3. Method according to one of claims 1 and 2, wherein the determination of at least one characteristic of the image comprises the determination of the eigenvalues of the image characterization matrix and the selection of the largest of said eigenvalues.
4. Method according to one of claims 1 to 3, wherein, in step c), a characteristic of the image is determined by calculating: _ ^2 + 22 in which, and correspond to eigenvalues of the image characterization matrix.
5. Method according to one of claims 1 to 4, further comprising determining at least one local orientation of the image from an eigenvalue of the structure tensor determined in step a) for a channel of the image and step d) comprises processing the local orientation by the object detection neural network.
6. Method according to one of claims 1 to 5, in which, in step a), the Riesz transform is a Riesz transform with Gaussian kernel.
7. A method according to one of claims 1 to 6, wherein the image is a color image which comprises three channels including a red channel, a blue channel and a green channel.
8. System for detecting objects in an image comprising a processor (2) configured to implement the method according to claims 1 to 7.
9. Computer program product comprising instructions which, when the program is executed by a computer, cause the latter to implement the method according to one of claims 1 to 7.