Method and device for target detection within radar images

Abstract

Method and Device for Target Detection in Radar Images The present invention relates to a method (100) for target detection in radar images, comprising the following steps: - creation (102) of a neural network, capable of receiving six distinct data channels as input and optimized for target detection, by combining: - a first U-shaped architecture (A1) of the U-Net type; - a second architecture (A2) comprising three compression stages (54, 56, 58) connected successively, said combination parallelizing the encoding portion of the first architecture (A1) and the second architecture (A2); - a training phase (104) of said neural network; - detection (108) of a target within a current radar image using, during an inference phase (106), said trained neural network. Figure for the abstract: Figure 3

Description

Title of the invention: Method and device for target detection within radar images

[0001] The present invention relates to a method for detecting targets within radar images.

[0002] The invention also relates to a computer program comprising software instructions which, when executed by a computer, implement such a process.

[0003] The present invention also relates to an electronic device for detecting targets within radar images.

[0004] The invention therefore lies in the field of image segmentation, and more specifically in the detection of target within radar images, in particular images from maritime surveillance radars, in order to provide, pixel by pixel, a binary classification representative of whether the pixel in question belongs to a target or not.

[0005] Traditionally, radar operators apply a predetermined constant false alarm rate TFAC to detect targets within radar images.

[0006] More specifically, such a TFAC process based on the application of a constant false alarm rate uses, in particular, a first preprocessing stage to enhance targets using spatial filtering. Then, local statistics are extracted to estimate the level of clutter (notably due to sea level) and thermal noise in the data. Finally, thresholding is applied to the radar images to retain only targets above the chosen threshold level. This thresholding can be adjusted to obtain a constant false alarm rate, below a value guaranteed by the radar manufacturer.

[0007] However, the TFAC method, based on the application of a constant false alarm rate, involves a significant number of parameters and requires readjustment each time the radar mode changes. Finally, in high sea states, the false alarm rate is also likely to be very high.

[0008] Furthermore, when seeking to determine, pixel by pixel, whether the pixel in question is part of a target, namely an artificial object in the environment such as the sea, or part of the seabed (i.e. sea clutter, thermal noise, etc.), the current target detection rate using the TFAC method based on the application of a false alarm rate remains improvable.

[0009] To provide, pixel by pixel, a more efficient binary classification representative of whether the pixel in question belongs to a target or not, the use of artificial intelligence for segmentation and object detection within radar images is beginning to be explored, but currently most existing neural networks are not yet efficient enough in terms of accuracy, inference speed, and embedding.

[0010] The aim of the invention is therefore to propose a solution using artificial intelligence to obtain more efficient radar image segmentation suitable for enabling the implementation of high-speed and high-precision target detection.

[0011] To this end, the invention relates to a method for detecting targets within radar images, the method being implemented by an electronic device, and comprising at least the following steps:

[0012] - creation of a neural network, capable of receiving six distinct data channels input and optimized for target detection, by combination:

[0013] - of a first U-shaped architecture of the U-Net type comprising symmetrically:

[0014] - an encoding part, designed to extract the characteristics of each radar image and comprising, linked successively, three encoding layers, ordered in ascending order, alternating with two subsampling layers,

[0015] - a decoding part suitable for reconstructing a high-resolution segmentation of each radar image, via three successive decoding layers ordered in descending order;

[0016] - of a second architecture comprising three interconnected compression stages successively,

[0017] said combination parallelizing the encoding part of the first architecture and the second architecture via:

[0018] - a first link connecting an output of the first sub-layer sampling of the encoding part at the input of the first compression stage;

[0019] - a second connection linking an output of the first compression stage to a input of a first summation element, inserted between the second coding layer and the second subsampling layer, and configured to sum said output of the first compression stage with the output of the second encoding layer and to provide a resulting output as input to the second subsampling layer;

[0020] - a third link connecting an output of the second sub- layer sampling of the encoding portion at one input of the second compression stage;

[0021] - a fourth connection linking the output of the second compression stage to a input of a second summation element, inserted between the third coding layer and a third subsampling layer, and configured to sum said output of the second compression stage with the output of the third layer encoding and to provide a resulting output to the input of the third subsampling layer;

[0022] - a fifth link connecting the output of the third compression stage to a input of a third summation element, inserted between the third subsampling layer and the third decoding layer, and configured to sum said output of the third compression stage with the output of the third subsampling layer and provide a resulting output as input to the third decoding layer;

[0023] - a training phase for said neural network;

[0024] - target detection within a current radar image using, during a phase of inference, said trained neural network capable of providing a boolean mask associating a first boolean value with each pixel of said current radar image in which a target element is detected and a second boolean value, distinct from said first value, with each pixel devoid of a target element.

[0025] Thus, the present invention proposes a solution based on the creation of a new neural network architecture, then its training, and its use once trained to provide target detection within radar images improved in terms of accuracy (particularly in the face of difficult sea states), inference speed, and portability, linked to the number of parameters.

[0026] To do this, the neural network created, then trained, then used, comprises the elements of a first U-shaped architecture of the U-Net type without superfluous layers, while maintaining adequate detection accuracy, only the coding elements of this first architecture being cleverly combined, for the specific needs according to the present invention of target detection, in parallel with a second image processing architecture of compression and subsampling, the decoding elements not being modified by parallel processing.

[0027] Thus, after each encoding layer of the first U-Net-type U-shaped architecture, the resulting latent space is retained and then optimally sampled using a compression stage of the second architecture (i.e., to decrease the resolution in the latent space), and then added to the next encoding layer, thereby increasing training and inference performance.

[0028] The specific combination (i.e., the specific arrangement of the neural network architecture created) according to the present invention, via the five judiciously placed links between the two architectures, and the gradual layer-by-layer subsampling in the first U-shaped U-Net architecture allows to achieve the desired performance, namely accuracy, inference speed and embedability.

[0029] According to other advantageous aspects of the invention, the method for detecting targets within radar images comprises one or more of the following features, taken individually or in any technically possible combination:

[0030] - said method further comprises, for said training phase, the preparation input data for said neural network comprising, for each training radar image in a training radar image database, the generation of six distinct data channels comprising:

[0031] - a first set of three channels of distinct data types, of which the data are normalized locally by radar image;

[0032] - a second set of three channels of distinct data types, of which the data are globally standardized for all radar images from the training base;

[0033] - for said first and second sets, the three data type channels distinct channels are respectively dedicated to signals in phase I, in quadrature phase Q, and in power P.

[0034] the distribution of each of the three channels of the first set being centered and reduced locally by radar training image of said radar training image base, the distribution of each of the three channels of the second set being centered and reduced with a mean and a standard deviation obtained by considering the set of radar images of the training base;

[0035] - the first encoding layer is convolutional and adapted for line detection horizontal within each radar image;

[0036] - the dimension of the convolution kernel of the first encoding layer is dimension (3,7);

[0037] - the first compression stage comprises two successive sets of layers:

[0038] - a first set of layers successively combining a layer of convolution, then a normalization layer called "Batch Norm", then a ReLu correction layer, the first set having sixteen input channels and sixteen output channels;

[0039] - a second set of layers successively combining a layer of convolution, then a normalization layer called "Batch Norm", then a ReLu correction layer, and modifying the sampling to go from sixteen input channels to thirty-two output channels;

[0040] - the second compression stage comprises successively:

[0041] - a first set of layers successively combining a layer of convolution, then a normalization layer called "Batch Norm", said first set having thirty-two input channels and sixteen output channels;

[0042] - an interpolation layer whose input is connected to the output of said first set of layers of said second compression stage;

[0043] - a summation element configured to sum the output of the first stage of compression with the output of the interpolation layer;

[0044] - connected to the output of said summation element, a second set of layers successively combining a convolution layer, then a normalization layer called "Batch Norm", then a ReLu correction layer, and modifying the sampling to go from sixteen input channels to thirty-two output channels;

[0045] - connected to the output of said second set of layers of said second stage of compression, a third set of layers successively combining a compression layer, a convolution layer, then a normalization layer called "Batch Norm", then a ReLu correction layer and modifying the sampling to go from thirty-two input channels to sixty-four output channels.

[0046] - the third compression stage comprises successively:

[0047] - connected to the output of said second set of layers of said second stage of compression, a first set of layers successively combining a compression layer, a convolution layer, then a normalization layer called "Batch Norm", then a ReLu correction layer and presenting thirty-two input channels and thirty-two output channels;

[0048] - connected to the output of said first set of layers of said third stage compression, a second set of layers successively combining a compression layer, a convolution layer, then a normalization layer called "Batch Norm", then a ReLu correction layer and modifying the sampling to go from thirty-two input channels to sixty-four output channels;

[0049] The invention also relates to a computer program comprising software instructions which, when executed by a computer, implement a method for detecting a target within radar images as defined above.

[0050] The invention also relates to an electronic target detection device within radar images comprising at least:

[0051] - a creation module configured to create a neural network, specific to receive six distinct input data channels and optimized for target detection, by combination:

[0052] - of a first U-shaped architecture of the U-Net type comprising symmetrically:

[0053] - an encoding part, designed to extract the characteristics of each radar image and comprising, linked successively, three encoding layers, ordered in ascending order, alternating with two subsampling layers,

[0054] - a decoding part suitable for reconstructing a high-resolution segmentation of each radar image, via three successive decoding layers ordered in descending order; and

[0055] - of a second architecture comprising three interconnected compression stages successively,

[0056] said combination parallelizing the encoding part of the first architecture and the second architecture via:

[0057] - a first link connecting an output of the first sub-layer sampling of the encoding part at the input of the first compression stage;

[0058] - a second connection linking an output of the first compression stage to a input of a first summation element, inserted between the second coding layer and the second subsampling layer, and configured to sum said output of the first compression stage with the output of the second encoding layer and to provide a resulting output as input to the second subsampling layer;

[0059] - a third link connecting an output of the second sub- layer sampling of the encoding portion at one input of the second compression stage;

[0060] - a fourth connection linking the output of the second compression stage to a input of a second summation element, inserted between the third coding layer and a third subsampling layer, and configured to sum said output of the second compression stage with the output of the third encoding layer and to provide a resulting output as input to the third subsampling layer;

[0061] - a fifth link connecting the output of the third compression stage to a input of a third summation element, inserted between the third subsampling layer and the third decoding layer, and configured to sum said output of the third compression stage with the output of the third subsampling layer and provide a resulting output as input to the third decoding layer;

[0062] - a training module configured to train said neural network;

[0063] - a detection module configured to detect at least one target within a current radar image using, during an inference phase, said trained neural network capable of providing a Boolean mask associating a first value a boolean value at each pixel of said current radar image in which a target element is detected and a second boolean value, distinct from said first value, at each pixel devoid of a target element.

[0064] According to another advantageous aspect of the invention, the electronic target detection device within radar images further comprises a data preparation module configured to prepare input data for said neural network, comprising, for each training radar image in a training radar image database, the generation of six distinct data channels comprising:

[0065] - a first set of three channels of distinct data types, of which the data are normalized locally by radar image;

[0066] - a second set of three channels of distinct data types, of which the data are globally normalized for all radar images from the training base.

[0067] The invention will become clearer upon reading the following description, given solely by way of non-limiting example, and made with reference to the drawings in which:

[0068] [Fig-1] [Fig.1] is a schematic representation of an electronic device target detection within radar images, according to the present invention.

[0069] [Fig.2] [Fig.2] is a schematic representation of an example of a neural network created according to the present invention.

[0070] [Fig.3] [Fig.3] is a flowchart of the main steps of a method for detecting a target within radar images, according to the present invention.

[0071] Fig. 1 illustrates an embodiment of an electronic target detection device 10 within radar images, according to the present invention.

[0072] The electronic device 10 comprises firstly a creation (i.e. generation) module 12 configured to create (i.e. generate) a neural network, illustrated and described hereafter in relation to [Fig. 2], designed to receive six distinct input data channels and optimized for target detection, by combination:

[0073] - of a first U-shaped architecture of the U-Net type comprising symmetrically:

[0074] - an encoding part, designed to extract the characteristics of each radar image and comprising, linked successively, three encoding layers, ordered in ascending order, alternating with two subsampling layers,

[0075] - a decoding part suitable for reconstructing a high-resolution segmentation of each radar image, via three successive decoding layers ordered in descending order; and

[0076] - of a second architecture comprising three interconnected compression stages successively,

[0077] said combination parallelizing the encoding part of the first architecture and the second architecture, called compression, via:

[0078] - a first link connecting an output of the first sub-layer sampling of the encoding part at the input of the first compression stage;

[0079] - a second connection linking an output of the first compression stage to a input of a first summation element, inserted between the second coding layer and the second subsampling layer, and configured to sum said output of the first compression stage with the output of the second encoding layer and to provide a resulting output as input to the second subsampling layer;

[0080] - a third link connecting an output of the second sub- layer sampling of the encoding portion at one input of the second compression stage;

[0081] - a fourth connection linking the output of the second compression stage to a input of a second summation element, inserted between the third coding layer and a third subsampling layer, and configured to sum said output of the second compression stage with the output of the third encoding layer and to provide a resulting output as input to the third subsampling layer;

[0082] - a fifth link connecting the output of the third compression stage to a input of a third summation element, inserted between the third subsampling layer and the third decoding layer, and configured to sum said output of the third compression stage with the output of the third subsampling layer and provide a resulting output as input to the third decoding layer;

[0083] In other words, said creation module 12 is capable of establishing the aforementioned connections specific to the present invention between known neural network architecture elements. Thus, said creation module 12, by establishing such connections, creates (i.e., generates) a specific arrangement of the overall architecture of the neural network dedicated, according to the present invention, to target detection within a radar image.

[0084] As illustrated in [Fig.1], the electronic target detection device 10 within radar images further comprises a training module 14 configured to train said neural network.

[0085] The electronic device 10 for target detection within radar images also includes a detection module 16 configured to detect at least one target within a current radar image by using, during an inference phase, said trained neural network suitable for providing a Boolean mask associating a first Boolean value with each pixel of said current radar image in which an element of target is detected and a second boolean value, distinct from said first value, is assigned to each pixel lacking a target element.

[0086] As an optional complement, as illustrated by the dotted line, the electronic target detection device 10 within radar images also includes a data preparation module 18 configured to prepare input data for said neural network comprising, for each training radar image from a training radar image database, the generation of six distinct data channels comprising:

[0087] - a first set of three channels of distinct data types, of which the data are normalized locally by radar image;

[0088] - a second set of three channels of distinct data types, of which the data are globally normalized for all radar images from the training base.

[0089] Indeed, according to a first variant, the training module 14, configured to train said neural network, directly receives said input data prepared beforehand by an external preparation module (not shown in [Fig.1]) to the electronic device 10 for target detection within radar images according to the present invention.

[0090] According to a second variant, illustrated by [Fig.1], the electronic device 10 for target detection within radar images is itself capable of implementing such data preparation at the beginning of the training phase, and this via the optional data preparation module 18.

[0091] In the example of [Fig.1], the electronic device 10 for detecting targets within radar images includes an information processing unit 20 formed for example of a memory 22 and a processor 24 associated with the memory 22.

[0092] In the example of [Fig. 1], the creation module, the training module, and the detection module, as well as the optional data preparation module, are each implemented as a software program, or a software component, executable by the processor 24. The memory 22 of the electronic target detection device within radar images is thus capable of storing creation software, training software, and detection software, as well as the optional data preparation software. The processor 24 is then capable of executing each of the following software programs: creation software, training software, and detection software, as well as the optional data preparation software.

[0093] In an alternative not shown, the creation module, the training module and the detection module, as well as the optional data preparation module, are each implemented as a programmable logic component, such as than an FPGA (from the English Field Programmable Gate Array), or even an integrated circuit, such as an ASIC (from the English Application Specifies Integrated Circuit).

[0094] When the electronic target detection device within radar images is implemented in the form of one or more software programs, i.e., in the form of a computer program, also called a computer program product, it is further capable of being stored on a computer-readable medium (not shown). The computer-readable medium is, for example, a medium capable of storing electronic instructions and being connected to a bus of a computer system. For example, the readable medium is an optical disc, a magneto-optical disc, ROM, RAM, any type of non-volatile memory (e.g., FLASH or NVRAM), or a magnetic card. A computer program comprising software instructions is then stored on the readable medium.

[0095] Fig. 2 is a schematic representation of an example of a neural network 30 created according to the present invention, and suitable for receiving six distinct data channels as input.

[0096] In general, a neural network comprises an ordered succession of layers of neurons, each of which takes its inputs from the outputs of the previous layer.

[0097] More precisely, each layer comprises neurons taking their inputs from the outputs of the neurons of the previous layer, or from the input variables for the first layer.

[0098] Alternatively, more complex neural network structures can be envisaged with a layer that can be linked to a layer further away than the immediately preceding layer.

[0099] Each neuron is also associated with an operation, that is to say a type of processing, to be carried out by said neuron within the corresponding processing layer.

[0100] Each layer is connected to the other layers by a plurality of synapses. A synaptic weight is associated with each synapse, and each synapse forms a link between two neurons. It is often a real number, which takes both positive and negative values. In some cases, the synaptic weight is a complex number.

[0101] Each neuron is capable of performing a weighted sum of the value(s) received from the neurons of the preceding layer, each value being multiplied by the respective synaptic weight of each synapse, or connection, between said neuron and the neurons of the preceding layer, then applying an activation function, typically a non-linear function, to said weighted sum, and delivering at the output of said neuron, in particular to the neurons of the next layer connected to it, the value resulting from the application of the activation function. The activation function This allows for the introduction of non-linearity in the processing performed by each neuron. The sigmoid function, the hyperbolic tangent function, and the Heaviside function are examples of activation functions.

[0102] As an optional complement, each neuron is also capable of applying, in addition, a multiplicative factor, also called bias, to the output of the activation function, and the value delivered at the output of said neuron is then the product of the bias value and the value from the activation function.

[0103] A convolutional neural network is also sometimes called a convolutional neural network or by the acronym CNN, which refers to the English term "Convolutional Neural Networks".

[0104] In a convolutional neural network, each neuron in the same layer has exactly the same connection pattern as its neighboring neurons, but at different input positions. The connection pattern is called the convolutional kernel or, more often, the "kernel" in reference to the corresponding English term.

[0105] A fully connected layer of neurons is a layer in which the neurons of said layer are each connected to all the neurons of the preceding layer.

[0106] Such a type of layer is more often referred to by the English term "fully connected", and sometimes designated by the name "dense layer".

[0107] According to the present invention, the neural network is cleverly created so as to be optimally adapted to target detection within radar images and includes first of all a first U-shaped U-Net type Ai architecture.

[0108] Traditionally, a U-Net neural network is a convolutional neural network model used for image segmentation and proposed by Ronneberger et al. in 2015 in the article entitled "U-Net: Convolutional Networks for Biomedical Image Segmentation". The U-Net is distinguished by its U-shaped architecture, which includes an encoding part to extract image features and a decoding part to reconstruct the segmentation in high resolution.

[0109] The present invention proposes to use a first U-shaped architecture Ai of the U-Net type, but devoid of any superfluous layer, the first architecture Ai comprising as illustrated by [Fig.2] comprising first of all an encoding part, suitable for extracting the characteristics of each radar image and comprising, linked successively, three encoding layers 31, 34, 36 ordered in an increasing order Eb E2 and E3 (from left to right according to the coding index going from one to three), these encoding layers being alternated with two subsampling layers 38 and 40 (from the English peeling) suitable for reducing the resolution by two.

[0110] In other words, in the encoding part of the first architecture Ab the first encoding layer 31 Ei (convolutional), with six distinct data input channels and sixteen output channels, is followed by a first subsampling layer 38 P reducing the resolution by half, for example to go from an input resolution of 512 pixels to an output resolution of 256 pixels, itself followed by the second encoding layer 34 E2 (convolutional) with sixteen input channels and thirty-two output channels, also followed in turn by a second subsampling layer 40 P reducing the resolution by half, for example to go from an input resolution of 256 pixels to an output resolution of 128 pixels, itself followed by the third encoding layer 36 E3 (convolutional) with thirty-two input channels and sixty-four output channels.

[0111] Symmetrically, the first architecture Ai includes a decoding part for reconstructing a high-resolution segmentation of each radar image, via three successive decoding layers 42, 44, 46 ordered in descending order (from left to right according to the decoding index going symmetrically from three to one), namely the D3 decoder in the decoding layer 42, then the D2 decoder in the decoding layer 44, then the Di decoder in the decoding layer 46.

[0112] As illustrated by [Fig.2], within the U-shaped form of the first Ai architecture, residual connections are established between two layers, respectively coding and decoding, symmetrical with respect to the axis of symmetry between the encoding and decoding parts, which allows the information of the input data to be retained, namely:

[0113] - a connection 48 between the encoder Ei of the encoding layer 31 at the end distal end of the encoding part, with respect to this axis of symmetry, and the decoder Di of the decoding layer 46 at the distal end of the decoding part, with respect to this axis of symmetry,

[0114] - a connection 50 between the encoder E2 of the encoding layer 34 at the center of the encoding section and the D2 decoder of the decoding layer 44 in the center of the decoding section, and

[0115] - a connection 52 between the encoder E3 of the encoding layer 36 at the end proximal, with respect to this axis of symmetry, and the D3 decoder of the decoding layer 42 at the proximal end of the decoding part, with respect to this axis of symmetry.

[0116] The neural network 30 created according to the present invention further comprises a second architecture A2, called compression, because it comprises three compression stages 54, 56, 58 connected successively.

[0117] It is to be noted that such compression stages 54, 56, 58 could be seen as extracted (i.e. as a selection of certain elements) from the DDRNet-type architecture introduced in 2021 by Hong et al. in the article entitled: “Deep Dual-resolution Networks for Real-time and Accurate Semantic Segmentation of Road Scenes” proposing the DDRNet neural network model designed specifically for object detection and segmentation in images in order to achieve an optimal balance between accuracy, inference speed and embedding capability on embedded systems.

[0118] However, the present invention relates specifically to the combination, with the first adapted U-Net type Ai architecture as previously described, of these adapted elements (as will be seen later) and selected to form said second architecture A2, said combination being based on the addition of specific links configured to parallelize the encoding part of the first Ai architecture and the second A2 architecture, known as the compression architecture, namely:

[0119] - a first link 60 connecting an output of the first sub-layer sampling 38 of the encoding part at the input of the first compression stage 54;

[0120] - a second link 62 connecting an output of the first compression stage 54 to an input of a first summation element 63, inserted between the second coding layer 34 and the second subsampling layer 40, and configured to sum said output of the first compression stage 54 with the output of the second encoding layer 34 and to provide a resulting output as input to the second subsampling layer 40;

[0121] - a third link 66 connecting an output of the second sub- layer sampling 40 of the encoding part to one input of the second compression stage 56;

[0122] - a fourth link 68 connecting the output of the second compression stage 56 to an input of a second summation element 70, interposed between the third coding layer 36 and a third subsampling layer 72, and configured to sum said output of the second compression stage 56 with the output of the third encoding layer 36 and to provide a resulting output as input to the third subsampling layer 72;

[0123] - a fifth link 74 connecting the output of the third compression stage 58 to an input of a third summation element 76, interposed between the third subsampling layer 72 and the third decoding layer 42, and configured to sum said output of the third compression stage 58 with the output of the third subsampling layer 72 and provide a resulting output as input to the third decoding layer 42.

[0124] Such an arrangement is specifically adapted to target detection that meets the criteria of accuracy, inference speed, and portability, and is not trivial in view of the aforementioned documents, which do not disclose such a combination of architectures, nor such links and their specific locations for achieving such an arrangement. In particular, it should be noted that in the DDRNet-type architecture introduced in 2021 by Hong et al., the parallelization between the lower compression architecture and a higher architecture, distinct from a U-Net-type architecture as in the present invention, begins after the second element of the upper DDRNet-type architecture and not, as according to the present invention, from the first encoding layer 31. This is more "natural" given the number of layers, to avoid the multi-scale compression processing, via the second architecture A2, spilling over onto the decoding layers.In other words, the present invention proposes to parallelize the three encoding layers of the first Ai architecture with the three compression stages of the second A2 architecture.

[0125] Optional complements associated with said neural network specially created according to the present invention via the creation module 12 for the needs of effective target detection within radar images are described below.

[0126] According to a first optional complement, the first convolutional encoding layer 31 is adapted for horizontal line detection within each radar image.

[0127] According to a variant of this optional complement, the dimension of the convolution kernel of the first encoding layer 31 is (3,7) (i.e., 3x7 in size). Such an odd-sized kernel (3,7) is indeed suitable for detecting targets with a rectangular aspect ratio (2,1) or (2,2). The advantage of odd-sized convolution kernels lies in the fact that this allows for a source image "centered" on the input pixels. Thus, the odd-sized rectangular kernel (3,7), according to the present invention, is suitable for radar targets, which are notably rectangular horizontal tracks in a sea clutter due to the radar capture method. Given such rectangular tracks, a rectangular kernel allows for better target identification after training.

[0128] Indeed, when a convolution is applied to an image, the convolution uses a kernel along the image dimension, for example 512 pixels by 512 pixels (i.e. 512x512).

[0129] This kernel can be of different sizes, for example in the U-Net initially described by Ronneberger et al., the starting kernel is of size 3x3 and consequently a padding of size one is used to fill the image and obtain an image of size 512 by 512 at the output of this layer.

[0130] Here, according to this optional complement, a 3x7 size kernel (rectangular targets) and a 1x3 size padding are used to fill the outgoing image of the convolution with one pixel in height for the dimension equal to three of the 3x7 size, and three pixels in width for the dimension equal to seven of the 3x7 size.

[0131] According to another optional supplement, the first compression stage 54 comprises two successive sets of layers, with firstly a first set of layers 78 combining successively a convolution layer, then a normalization layer called "Batch Norm", then a ReLu correction layer. Said first set has sixteen input channels and sixteen output channels and is notably called "CBNR 16,16", where "C" represents the convolution layer, "BN" represents the normalization layer called "Batch Norm" and "R" represents the ReLu layer, "16" representing, within "CBNR 16,16", the number of input channels and "16" representing the number of output channels.

[0132] Following this first set 78, the first compression stage 54 comprises a second set of layers 80 combining successively a convolution layer, then a normalization layer called "Batch Norm", then a ReLu correction layer, and modifying the sampling to go from sixteen input channels to thirty-two output channels. Said second set is notably called "CBNR 16, 32", where "C" represents the convolution layer, "BN" represents the normalization layer called "Batch Norm" and "R" represents the ReLu layer, "16" representing, within "CBNR 16, 32", the number of input channels and "32" representing the number of output channels.

[0133] According to another optional supplement, the second compression stage 56 successively comprises first of all a first set of layers 82 combining successively a convolution layer, then a normalization layer called "Batch Norm", said first set having thirty-two input channels and sixteen output channels, and called "CBN 32,16", of which "C" represents the convolution layer, "BN" represents the normalization layer called "Batch Norm", "32" representing, within "CBN 32,16", the number of input channels and "16" representing the number of output channels.

[0134] Following this first set 82, the second compression stage 56 includes an interpolation layer 84 called "Interpol" whose input is connected to the output of said first set of layers 82 of said second compression stage 56, and suitable for increasing the resolution again, for example, to go from an input resolution of 128 pixels to an output resolution of 256 pixels.

[0135] Successive to this interpolation layer 84, the second compression stage 56 includes a summation element 86 configured to sum the output of the first compression stage 54 with the output of the interpolation layer 84.

[0136] Connected, at the input, to the output of said summation element 86, the second compression stage 56 comprises a second set of layers 88 successively combining a convolution layer, then a normalization layer called "Batch Norm", then a ReLu correction layer, and modifying the sampling to go from sixteen input channels to thirty-two output channels. This second set of layers 88 of the second compression stage 56 is also notably called "CBNR 16, 32", where "C" represents the convolution layer, "BN" represents the normalization layer called "Batch Norm" and "R" represents the ReLu layer, "16" representing, within "CBNR 16, 32", the number of input channels and "32" representing the number of output channels.

[0137] Finally, according to this optional supplement, the second compression stage 56 also includes, connected at the input to the output of said second set of layers 88, a third set of layers 90 combining successively a compression layer reducing the resolution by two, for example to go from an input resolution of 256 pixels to an output resolution of 128 pixels (i.e. compatible with the third E3 encoding layer 36 via a parameter called "stride" which denotes the number of pixels by which the window moves after each operation), a convolution layer, then a normalization layer called "Batch Norm", then a ReLu correction layer and modifying the sampling to go from thirty-two input channels to sixty-four output channels.This third set of layers 90 of the second compression stage 56 is notably called "cCBNR 32, 64", where "c" represents the compression layer, "C" represents the convolution layer, "BN" represents the normalization layer called "Batch Norm" and "R" represents the Relu layer, "32" representing, within "cCBNR 32, 64", the number of input channels and "64" representing the number of output channels.

[0138] According to another optional supplement, the third compression stage 58 comprises successively, connected to the output of said second set of layers 88 of said second compression stage 56, firstly a first set of layers 92 combining successively a compression layer, a convolution layer, then a normalization layer called "Batch Norm", then a ReLu correction layer and having thirty-two input channels and thirty-two output channels.

[0139] This first set of layers 92 of the third compression stage 58 is notably called "cCBNR 32, 32", where "c" represents the compression layer reducing the resolution by half, for example, to go from an input resolution of 256 pixels to an output resolution of 128 pixels (via a parameter called "stride" which denotes the number of pixels per (in which the window moves after each operation), "C" represents the convolution layer, "BN" represents the normalization layer called "Batch Norm" and "R" represents the Relu layer, "32" representing, within "cCBNR 32, 32", the number of input channels and "32" representing the number of output channels.

[0140] According to this optional complement, connected to the output of said first set of layers 92 of said third compression stage 58, a second set of layers 94 successively combining a compression layer "c" reducing the resolution by half, for example, to go from an input resolution of 128 pixels to an output resolution of sixty-four pixels (via a parameter called "stride" which denotes the number of pixels by which the window moves after each operation), a convolution layer "C", then a normalization layer called "Batch Norm" or "BN", then a ReLu correction layer "R" and modifying the sampling to go from thirty-two input channels to sixty-four output channels, this second set of layers 94 being called "cCBNR 32, 64", "32" representing, within "cCBNR 32, 64", the number of input channels and "64" representing the number of channels of exit .

[0141] Note that such an arrangement of the third compression stage 58 advantageously allows, via an additive compression feedback, via the summation element 76, to avoid the deep aggregation pyramid pooling module called "DAPPM" (from the English Deep Aggregation Pyramid Pooling Module) of the DDRNet type architecture introduced in 2021 by Hong et al.

[0142] It should be noted that, in general, the normalization layers called "Batch Norm" are activated only during the training phase of the neural network 30 and are inactive by default during the inference phase.

[0143] The first Ai architecture makes it possible to achieve the desired performance in terms of precision, while the second A2 architecture makes it possible to reduce the amount of information needed for target detection by reducing the image resolution and thus provides the advantages of multi-scale segmentation.

[0144] An example of the operation of said electronic target detection device 10 within radar images is now described below in relation to [Fig.3].

[0145] More specifically, according to the embodiment of the present invention illustrated by [Fig. 3], the method 100 comprises a first step 102 of CRE creation of a neural network N, suitable for receiving six distinct data channels as input and optimized for target detection, by combining on the one hand a first U-shaped architecture of the U-Net type comprising symmetrically an encoding part, suitable for extracting the characteristics of each radar image and comprising, connected successively, three encoding layers, ordered in ascending order, alternating with two subsampling layers, and a decoding part specific to reconstruct a high-resolution segmentation of each radar image, via three successive decoding layers ordered in descending order, and on the other hand a second architecture comprising three successively linked compression stages.

[0146] As previously stated, during said creation step 102, said combination cleverly parallelizes the encoding part of the first architecture and the second architecture via:

[0147] - a first link connecting an output of the first sub-layer sampling of the encoding part at the input of the first compression stage;

[0148] - a second link connecting an output of the first compression stage to a input of a first summation element, inserted between the second coding layer and the second subsampling layer, and configured to sum said output of the first compression stage with the output of the second encoding layer and to provide a resulting output as input to the second subsampling layer;

[0149] - a third link connecting an output of the second sub- layer sampling of the encoding portion at one input of the second compression stage;

[0150] - a fourth link connecting the output of the second compression stage to a input of a second summation element, inserted between the third coding layer and a third subsampling layer, and configured to sum said output of the second compression stage with the output of the third encoding layer and to provide a resulting output as input to the third subsampling layer;

[0151] - a fifth link connecting the output of the third compression stage to a input of a third summation element, inserted between the third subsampling layer and the third decoding layer, and configured to sum said output of the third compression stage with the output of the third subsampling layer and provide a resulting output as input to the third decoding layer.

[0152] Then, the process 100 includes a training phase 104 P_E of said neural network N, created previously in step 102.

[0153] Indeed, for its implementation, the neural network N must first be trained, that is to say confronted with massive data representative of the tactical situations that the radar may encounter.

[0154] Then, generally, the method 100 according to the present invention comprises, during an inference phase 106, a target detection step 108 within a current radar image IRC using said trained neural network NE from of the previous step 104 and suitable for providing a Boolean mask M associating a first Boolean value with each pixel of said current radar image IRC in which a target element is detected and a second Boolean value, distinct from said first value, with each pixel devoid of a target element.

[0155] Two variants of the implementation of the training phase are described below.

[0156] Indeed, according to a first variant, during this training phase 104, the training module 14, configured to train said neural network, directly receives said input data E' prepared beforehand, according to a step external to the process and not shown in [Fig.3].

[0157] According to a second variant, illustrated by [Fig.3], the training phase 104 includes a data preparation step 110, implemented at the beginning of the training phase.

[0158] This preparation 110 of input data for said neural network comprises, for each radar image (in particular two-dimensional 2D distance / recurrence after pulse compression) used for training a training radar image database BD, the generation of six distinct data channels comprising:

[0159] - a first set of three channels of distinct data types, of which the data are normalized locally by radar image;

[0160] - a second set of three channels of distinct data types, of which the data are globally normalized for all radar images from the training base.

[0161] More specifically, for said first and second sets, the three distinct data type channels are respectively dedicated to I-phase, Q-phase quadrature, and P-power signals (with P = I2+Q2), the distribution of each of the three channels of the first set being locally centered and reduced by training radar image of said training radar image base, the distribution of each of the three channels of the second set being centered and reduced with a mean and standard deviation obtained by considering the set of radar images of the training base.

[0162] Indeed, according to this optional supplement, this training requires data standardization, first locally, with the distribution of each channel being centered and reduced, and then globally, with each channel being centered and reduced using the mean and standard deviation observed across the entire dataset formed by all the training radar images in the training radar image database BD. The six channels, I, Q, P, standardized locally, and I, Q, and P, standardized globally, are provided as training input data for the neural network N, and are compared to the Boolean mask representing the ground truth, the Boolean mask classifying each pixel as "target" or "not a target".

[0163] Note that, according to an optional aspect, the trained NE neural network is then suitable for export in a standard format (for example ONNX©) and then implemented directly on the radar, or in post-processing.

[0164] It should be noted that, in inference, according to this optional complement, the data of the current IRC radar image are prepared in a similar way so as to provide the trained neural network with six distinct data channels from said current IRC radar image as input. More precisely, for said current IRC radar image, a first set of three channels of distinct data types whose data are locally normalized is obtained, and a second set of three channels of distinct data types whose data are globally normalized is obtained by using said current IRC radar image (its local values) with the data from all the radar images in the training database.

[0165] A person skilled in the art will understand that the invention is not limited to the embodiments described, nor to the particular examples of the description, the embodiments and variants mentioned above being capable of being combined with each other to generate new embodiments of the invention.

[0166] The present invention thus makes it possible to create and use a neural network fully adapted to target detection by using a second compression / subsampling architecture which increases the efficiency of the first U-Net type architecture which is already lighter compared to the article by Ronneberger et al., and which improves its inference speed.

[0167] This yields a target detection solution within at least one radar image that is both fast (i.e., high-speed with a massive volume of input data), capable of processing up to forty images per second depending on the target type; portable and energy-efficient, requiring fewer than one million parameters (or even fewer than 200,000 parameters, for example, 142,001 parameters), thus reducing complexity and limiting operational overhead; and accurate, as more than 70 to 75% of objects have been detected in all sea states using this solution, which is more effective in terms of false alarms than the TFAC method described above. In other words, the false alarm rate is reduced with the present invention.

[0168] Note that the present solution is presented here as dedicated to the detection of targets within radar images, but that the present invention could be applied similarly in other areas of image processing.

Claims

1. Demands Method (100) for detecting targets within radar images, the method being implemented by an electronic device, and comprising at least the following steps: - creation (102) of a neural network, capable of receiving six distinct data channels as input and optimized for target detection, by combination: - of a first architecture (U-shaped AJ of the U-Net type comprising symmetrically: - an encoding section, designed to extract the features of each radar image and comprising, successively linked, three encoding layers (31, 34, 36), ordered in ascending order, alternating with two subsampling layers (38, 40); - a decoding section designed to reconstruct a high-resolution segmentation of each radar image, via three successive decoding layers (42, 44, 46) ordered in descending order; and - a second architecture (A2) comprising three compression stages (54, 56, 58) linked successively, said combination parallelizing the encoding section of the first architecture (Ai) and the second architecture (A2) via: - a first link (60) connecting an output of the first subsampling layer (38) of the encoding part to the input of the first compression stage (54); - a second link (62) connecting an output of the first compression stage (54) to an input of a first summation element (63), interposed between the second coding layer (34) and the second subsampling layer (40), and configured to sum said output of the first compression stage (54) with the output of the second encoding layer (34) and to provide a resulting output as input to the second subsampling layer (40); - a third link (66) connecting an output of the second subsampling layer (40) of the encoding part to an input of the second compression stage (56); - a fourth link (68) connecting the output of the second compression stage (56) to an input of a second summation element (70), interposed between the third coding layer (36) and a

2.

3. third subsampling layer (72), and configured to sum said output of the second compression stage (56) with the output of the third encoding layer (36) and to provide a resulting output as input to the third subsampling layer (72); - a fifth link (74) connecting the output of the third compression stage (58) to an input of a third summation element (76), interposed between the third subsampling layer (72) and the third decoding layer (42), and configured to sum said output of the third compression stage (58) with the output of the third subsampling layer (72) and provide a resulting output as input to the third decoding layer (42); - a training phase (104) of said neural network; - detection (108) of target within a current radar image using, during an inference phase (106), said trained neural network capable of providing a boolean mask associating a first boolean value with each pixel of said current radar image in which a target element is detected and a second boolean value, distinct from said first value, with each pixel devoid of a target element. Method (80) according to claim 1, further comprising, for said training phase (104), the preparation (110) of input data of said neural network comprising, for each training radar image of a training radar image database, the generation of six distinct data channels comprising: - a first set of three channels of distinct data types whose data are locally normalized per radar image; - a second set of three channels of distinct data types whose data are globally normalized for all radar images of the training base. Method (80) according to claim 2, wherein, for said first and second assemblies, the three distinct data type channels are respectively dedicated to the I-phase, Q-phase quadrature, and P-power signals, the distribution of each of the three channels of the first assembly being locally centered and reduced by radar training image of said radar training image base, the distribution of each of the three channels of the second assembly being centered and reduced with an average and a standard deviation obtained by considering all the radar images from the training base.

4. A method (80) according to any one of the preceding claims, wherein the first encoding layer (31) is convolutional and adapted for horizontal line detection within each radar image.

5. Method (80) according to claim 4, wherein the dimension of the convolution kernel of the first encoding layer is dimension (3,7).

6. A method (80) according to any one of the preceding claims, wherein the first compression stage (54) comprises two successive sets of layers: - a first set of layers (78) successively combining a convolution layer, then a normalization layer called "Batch Norm", then a ReLu correction layer, said first set having sixteen input channels and sixteen output channels; - a second set of layers (80) successively combining a convolution layer, then a normalization layer called "Batch Norm", then a ReLu correction layer, and modifying the sampling to go from sixteen input channels to thirty-two output channels.

7. A method (80) according to any one of the preceding claims, wherein the second compression stage (56) successively comprises: - a first set of layers (82) successively combining a convolution layer and then a normalization layer called "Batch Norm", said first set having thirty-two input channels and sixteen output channels; - an interpolation layer (84) whose input is connected to the output of said first set of layers of said second compression stage; - a summation element (86) configured to sum the output of the first compression stage with the output of the interpolation layer; - connected to the output of said summation element (86), a second set of layers (88) successively combining a convolution layer and then a normalization layer called " "Batch Norm", then a ReLu correction layer, and modifying the sampling to go from sixteen input channels to thirty-two output channels; - connected to the output of said second set of layers (88) of said second compression stage (56), a third set of layers (90) successively combining a compression layer, a convolution layer, then a normalization layer called "Batch Norm", then a ReLu correction layer and modifying the sampling to go from thirty-two input channels to sixty-four output channels.

8. A method (80) according to any one of the preceding claims, wherein the third compression stage (58) comprises successively: - connected to the output of said second set of layers (88) of said second compression stage (56), a first set of layers (92) successively combining a compression layer, a convolution layer, then a normalization layer called "Batch Norm", then a ReLu correction layer and having thirty-two input channels and thirty-two output channels; - connected to the output of said first set of layers (92) of said third compression stage (58), a second set of layers (94) successively combining a compression layer, a convolution layer, then a normalization layer called "Batch Norm", then a ReLu correction layer and modifying the sampling to go from thirty-two input channels to sixty-four output channels.

9. Electronic target detection device (10) within radar images comprising at least: - a creation module (12) configured to create a neural network, suitable for receiving six distinct data channels as input and optimized for target detection, by combining: - a first architecture (U-shaped AJ of the U-Net type) comprising symmetrically: - an encoding part, suitable for extracting the features of each radar image and comprising, linked successively, three encoding layers (31, 34, 36), ordered in ascending order, alternating with two subsampling layers (38, 40), - a decoding section designed to reconstruct a high-resolution segmentation of each radar image, via three successive decoding layers (42, 44, 46) ordered in descending order; and - a second architecture (A2) comprising three successively connected compression stages (54, 56, 58), said combination parallelizing the encoding section of the first architecture (Ai) and the second architecture (A2) via: - a first link (60) connecting an output of the first subsampling layer (38) of the encoding part to the input of the first compression stage (54); - a second link (62) connecting an output of the first compression stage (54) to an input of a first summation element (63), interposed between the second coding layer (34) and the second subsampling layer (40), and configured to sum said output of the first compression stage (54) with the output of the second encoding layer (34) and to provide a resulting output as input to the second subsampling layer (40); - a third link (66) connecting an output of the second subsampling layer (40) of the encoding part to an input of the second compression stage (56); - a fourth link (68) connecting the output of the second compression stage (56) to an input of a second summation element (70), interposed between the third coding layer (36) and a third subsampling layer (72), and configured to sum said output of the second compression stage (56) with the output of the third encoding layer (36) and to provide a resulting output as input to the third subsampling layer (72); - a fifth link (74) connecting the output of the third compression stage (58) to an input of a third summation element (76), interposed between the third subsampling layer (72) and the third decoding layer (42), and configured to sum said output of the third compression stage (58) with the output of the third subsampling layer (72) and provide a resulting output as input to the third decoding layer (42); - a training module (14) configured to train said neural network;

10. - a detection module (16) configured to detect at least one target within a current radar image by using, during an inference phase, said trained neural network adapted to provide a Boolean mask associating a first Boolean value with each pixel of said current radar image in which a target element is detected and a second Boolean value, distinct from said first value, with each pixel lacking a target element. An electronic device (10) according to claim 9, further comprising a data preparation module (18) configured to prepare input data for said neural network comprising, for each training radar image in a training radar image database, the generation of six distinct data channels comprising: - a first set of three channels of distinct data types whose data are locally normalized by radar image; - a second set of three channels of distinct data types whose data are globally normalized for all radar images of the training base.

Images