A camouflaged target detection method and system based on learning neural network

Abstract

The present invention discloses a camouflaged target detection method and system based on a learning neural network in the field of camouflaged target detection. The method comprises the following steps: collecting a detection image and inputting it into a pre-trained target search model to obtain a multi-layer camouflaged target representation arranged in sequence from a low level to a high level; extracting camouflaged representation fragments from each layer of camouflaged target representation and inputting them into a pre-trained target recognition model to obtain a plurality of camouflaged recognition features of a corresponding level; convolving each camouflaged recognition feature to obtain a camouflaged recognition level image, and superimposing each camouflaged recognition level image to obtain a camouflaged recognition image; the present invention amplifies the receptive field of the extracted camouflaged representation f through a receptive field network, extracts camouflaged representation fragments at multiple scales and inputs them into a target recognition model, reassembles the camouflaged representation, captures global and local context detail information, can better recognize edge texture structures, and thus obtains a higher recognition accuracy.

Description

Technical Field

[0001] The present invention belongs to the field of camouflaged target detection, and in particular relates to a camouflaged target detection method and system based on a learning neural network. Background Art

[0002] Disguised target recognition is a key technology for many applications, including computer vision (which can be used for search and rescue work, or to find rare species), medical image analysis (such as polyp segmentation and pneumonia segmentation), agriculture (such as locust invasion monitoring), and art (for realistic image fusion or artistic entertainment). In general, camouflaged target recognition remains a challenging problem for two main reasons: first, the camouflaged object has a high similarity with its background, making it difficult to extract the camouflaged target representation; second, the contextual details of the camouflaged target are difficult to effectively enhance. Many organisms in nature have evolved to be confused with their surroundings, making them even more difficult to detect.

[0003] In the detection of disguised targets, due to the high similarity between the environment and the target, the extracted target representation is not accurate, and the detailed information of the context is not deeply mined. In addition, the detection accuracy of disguised targets is also highly dependent on the threshold. The target detection method in the existing technology cannot achieve good performance, and how to choose a suitable threshold is also a difficult problem. Therefore, the research on disguised target recognition is of great significance and has attracted widespread attention in recent years. Summary of the invention

[0004] The purpose of the present invention is to provide a camouflaged target detection method and system based on a learning neural network, in which the camouflaged target representation is input into the learning neural network in a hierarchical manner through progressive segmentation, so as to improve the detection accuracy of the camouflaged target.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is:

[0006] A first aspect of the present invention provides a camouflaged target detection method based on a learning neural network, comprising:

[0007] Collect the detection image and input it into the pre-trained target search model to obtain a multi-layer camouflage target representation that is set in sequence from low level to high level;

[0008] Extract camouflage representation fragments from each layer of camouflage target representation and input them into a pre-trained target recognition model to obtain multiple camouflage recognition features of corresponding levels;

[0009] Convolving each camouflage recognition feature to obtain a camouflage recognition level image, and superimposing each camouflage recognition level image to obtain a camouflage recognition image;

[0010] The training process of the target search model and the target recognition model includes:

[0011] Collect training images containing camouflaged targets to obtain a training data set;

[0012] The target search model and target recognition model will be trained through the training data set, and the structured loss value S of camouflage recognition will be calculated during the training process. The parameters of the target search model and target recognition model will be iteratively adjusted according to the structured loss value S to obtain the target search model and target recognition model that meet the set requirements.

[0013] Preferably, the method of collecting and detecting images and inputting them into a pre-trained target search model to obtain a multi-layer camouflage target representation arranged in sequence from low level to high level includes:

[0014] The detection image is input into the Res2Net50 residual network in the target search model, and the four-layer camouflage representation f is output in order from low level to high level;

[0015] The local detail information of each layer of camouflage representation f is amplified through the receptive field network to obtain camouflage target representation fr2, camouflage target representation fr3, camouflage target representation fr4 and camouflage target representation fr5 respectively.

[0016] Preferably, the method of extracting camouflage representation fragments from each layer of camouflage target representation includes:

[0017] Extract four slices of the disguised target representation fr5 as the disguised representation slice fr′5; extract two slices of the disguised target representation fr4 as the disguised representation slice fr′4; extract the entirety of the disguised target representation fr3 as the disguised representation slice fr′3; extract the entirety of the disguised target representation fr2 as the disguised representation slice fr′2.

[0018] Preferably, the method of inputting the camouflage representation fragments into a pre-trained target recognition model to obtain a plurality of camouflage recognition features of corresponding levels includes:

[0019] Input the camouflage representation fragment fr′5, the camouflage representation fragment fr′4, the camouflage representation fragment fr′3 and the camouflage representation fragment fr′2 into the target recognition model in order from low level to high level;

[0020] Perform wavelet attention transform on the lowest level camouflage representation fragment fr′5 to extract camouflage recognition feature r5;

[0021] Slice the next level of camouflage representation rf n and the camouflage identification feature r extracted at the previous level n+1 After combining, the camouflage recognition features r are obtained in turn through wavelet attention transform n .

[0022] Preferably, the next level of camouflage representation fragment rf n and the camouflage identification feature r extracted at the previous level n+1 After combining, the methods of obtaining camouflage identification features in sequence through wavelet attention transform include:

[0023] Slice the next level of camouflage representation rf n and the camouflage identification feature r extracted at the previous level n+1 After combining, convolution is performed to obtain low-frequency component LL, high-frequency component LH, high-frequency component HL and high-frequency component HH;

[0024] The low-frequency component LL captures the camouflaged segmentation information and contour information of the camouflaged target; the low-frequency component LL is enhanced by the attention mechanism AM and the instance normalized residual block ResIN, and the camouflaged segmentation information and contour information are superimposed to obtain the low-frequency component camouflage feature r LL ;

[0025] The high-frequency component LH, high-frequency component HL and high-frequency component HH are enhanced by the normalized residual block ResB and the attention mechanism AM respectively to obtain the high-frequency component camouflage feature r LH , high frequency component camouflage feature r HL and high frequency component camouflage feature r HH ;

[0026] The low frequency component disguises the feature r LL , high frequency component camouflage feature r LH , high frequency component camouflage feature r HL and high frequency component camouflage feature r HH Perform transposed convolution TC to obtain the camouflage recognition feature r k .

[0027] Preferably, the structural loss value S of camouflage recognition during the training process is calculated using the following formula:

[0028] S=α*S o +(1-α)*S r

[0029]

[0030] S o =μ*O FG +(1-μ)*O BG

[0031] Among them, S o It is expressed as an object-oriented structural similarity measure; S r It is represented as a region-oriented structural similarity measure; α is represented as a weight factor; W krepresents the weight of the kth camouflage representation slice; K represents the total number of camouflage representation slices; k represents the sequence number of the camouflage representation slice; μ is the proportion of the foreground area in the image; O BG Denoted as background measure; O FG Denoted as a foreground metric.

[0032] Preferably, the background metric O is calculated BG and the foreground metric O FG , the expression formula is:

[0033]

[0034]

[0035]

[0036]

[0037] Among them, D FG It is expressed as the difference between the foreground area of ​​the camouflage prediction map and the camouflage representation fragment, x FG Represents the probability value of the foreground area of ​​the camouflaged target prediction map, y FG They represent the probability values ​​of the foreground area of ​​the camouflage representation fragment, Expressed as the probability value x FG The mean of Expressed as probability values ​​y FG The mean of Expressed as the probability value x FG The variance, D BG It is expressed as the difference between the background area of ​​the camouflage prediction image and the camouflage representation fragment, x BG Represents the probability value of the background area of ​​the camouflaged target prediction map, y BG They represent the probability values ​​of the background area of ​​the camouflage representation fragment respectively, Expressed as the probability value x BG The mean of Expressed as probability values ​​y BG The mean of Expressed as the probability value x BG The variance of .

[0038] A second aspect of the present invention provides a camouflaged target detection system based on a learning neural network, comprising:

[0039] An acquisition module, which acquires detection images;

[0040] A target search module is used to input the collected detection image into a pre-trained target search model to obtain a multi-layer camouflage target representation arranged in sequence from low level to high level;

[0041] The target recognition module extracts camouflage representation fragments from each layer of camouflage target representation and inputs them into a pre-trained target recognition model to obtain multiple camouflage recognition features of corresponding levels;

[0042] A camouflage recognition image generation module convolves each camouflage recognition feature to obtain a camouflage recognition level image, and superimposes each camouflage recognition level image to obtain a camouflage recognition image;

[0043] The training module is used to train the target search model and the target recognition model.

[0044] Preferably, the target search model includes sequentially setting a feature extraction layer Stage1, a feature extraction layer Stage2, a feature extraction layer Stage3, a feature extraction layer Stage4 and a feature extraction layer Stage5; the feature extraction layer Stage2, the feature extraction layer Stage3, the feature extraction layer Stage4 and the feature extraction layer Stage5 sequentially extract camouflage representation f from the collected image; the feature extraction layer Stage2, the feature extraction layer Stage3, the feature extraction layer Stage4 and the feature extraction layer Stage5 are respectively connected to a receptive field network for amplifying local detail information of the camouflage representation f;

[0045] The target recognition model includes four wavelet attention units for wavelet attention transformation; the four wavelet attention units are connected to four receptive field networks in a one-to-one correspondence.

[0046] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of the camouflaged target detection method when executed by a processor.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] (1) The detection image collected by the present invention is input into a pre-trained target search model to obtain a multi-layer camouflage target representation arranged in sequence from low level to high level; a receptive field network is provided in the target search model, and the receptive field of the extracted camouflage representation f is amplified by the receptive field network, which is beneficial to improving the accuracy of the recognition rate.

[0049] (2) The present invention extracts camouflage representation fragments from each layer of camouflage target representation and inputs them into a pre-trained target recognition model to obtain multiple camouflage recognition features of corresponding levels; by extracting camouflage representation fragments at multiple scales and inputting them into the target recognition model, the camouflage representation is reassembled to capture global and local contextual detail information, which can better identify edge texture structures and thus obtain a higher recognition accuracy.

[0050] (3) The present invention convolves each camouflage recognition feature to obtain a camouflage recognition level image, and superimposes each camouflage recognition level image to obtain a camouflage recognition image; and effectively detects the camouflaged target through a target recognition model based on a wavelet learning network, thereby improving the accuracy of the recognition rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 is a flow chart of a fatigue detection method based on gait information provided by an embodiment of the present invention;

[0052] Figure 2 is a structural diagram of a receptive field network provided by an embodiment of the present invention;

[0053] Figure 3 It is a visual comparison diagram of identifying camouflaged targets by various methods provided by the embodiments of the present invention;

[0054] Figure 4 is a structural diagram of a wavelet attention unit provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0055] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0056] Embodiment 1

[0057] like Figure 1 As shown, the first aspect of the present invention provides a camouflaged target detection method based on a learning neural network, comprising:

[0058] The method of collecting and detecting images and inputting them into a pre-trained target search model to obtain a multi-layer camouflage target representation arranged in sequence from low level to high level includes:

[0059] Input the detection image into the Res2Net50 residual network in the target search model, wherein the Res2Net50 residual network includes a feature extraction layer Stage1, a feature extraction layer Stage2, a feature extraction layer Stage3, a feature extraction layer Stage4 and a feature extraction layer Stage5, covering a diversified feature pyramid from high resolution and weak semantics to low resolution and strong semantics, and the feature extraction layer Stage2, the feature extraction layer Stage3, the feature extraction layer Stage4 and the feature extraction layer Stage5 output four layers of camouflage representation f arranged in sequence from low level to high level;

[0060] The local detail information of each layer of camouflage representation f is amplified through the receptive field network to obtain camouflage target representation fr2, camouflage target representation fr3, camouflage target representation fr4 and camouflage target representation fr5 respectively. The receptive field module is a multi-branch convolution block, and its internal structure can be divided into the following parts: multi-branch convolution layer with different kernels and trailing dilated convolution layer. The branch convolution layer is responsible for simulating the multi-scale receptive field function, and the trailing dilated convolution layer reproduces the relationship between the receptive field size and eccentricity in the human visual system.

[0061] Among them, Figure 2 As shown in the figure, the receptive field module includes five branches. The first branch is convolution 1×1; the second branch is sequentially configured with convolution 1×1, convolution 1×3, convolution 3×1 and dilated convolution ×3; the third branch is sequentially configured with convolution 1×1, convolution 1×5, convolution 5×1 and dilated convolution ×5; the second branch is sequentially configured with convolution 1×1, convolution 1×7, convolution 7×1 and dilated convolution ×7; after the first four branches are connected in series, through the 1×1 convolution layer operation, its channels are reduced to the predetermined number of channels. Finally, the fifth branch is added and the whole is input into the activation function to obtain the enhanced features.

[0062] The method of extracting camouflage representation fragments from each layer of camouflage target representation includes:

[0063] Extract four slices of the disguised target representation fr5 as the disguised representation slice fr′5; extract two slices of the disguised target representation fr4 as the disguised representation slice fr′4; extract the entirety of the disguised target representation fr3 as the disguised representation slice fr′3; extract the entirety of the disguised target representation fr2 as the disguised representation slice fr′2.

[0064] The method of inputting the camouflage representation fragments into a pre-trained target recognition model to obtain multiple camouflage recognition features of corresponding levels includes:

[0065] Input the camouflage representation fragment fr′5, the camouflage representation fragment fr′4, the camouflage representation fragment fr′3 and the camouflage representation fragment fr′2 into the target recognition model in order from low level to high level;

[0066] Perform wavelet attention transform on the lowest level camouflage representation fragment fr′5 to extract camouflage recognition feature r5;

[0067] Slice the next level of camouflage representation rf n and the camouflage identification feature r extracted at the previous level n+1 After the combination, the camouflage recognition features are obtained in turn through wavelet attention transformation. The specific methods include:

[0068] The present invention uses wavelet transform to decompose the features in the frequency domain. The frequency subbands are reconstructed by addition and transposed convolution. For the pixels in 2×2 blocks (each block is represented as a, b, c and d), the calculation process of the two-dimensional Haar wavelet coefficients is defined as:

[0069]

[0070] Among them, A represents low-frequency information. B, C, and D represent high frequencies in the horizontal, vertical, and diagonal directions, respectively. The height and width of the decomposed subbands are half the size of the original image. Then the high-frequency subbands and low-frequency subbands are enhanced respectively. Among them, the high-frequency information mainly contains texture and edge information. The high-frequency information is extracted and connected to the decoder to retain edge-like information. In addition, a residual block with an attention module is used to enhance the high-frequency subband. Specifically, the attention module consists of channel attention and spatial attention. Since high-frequency information is sparse, the maximum pooling is used instead of the average pooling to emphasize important channels. Then this module makes the wavelet learning network pay more attention to the spatial information features. For low-frequency information, we propose a new low-frequency enhancement branch with normalization and attention mechanism. The image is converted to an arbitrary style through adaptive instance normalization. This layer aligns the mean and variance of the disguised segmentation information (content representation) with the mean and variance of the contour information (style representation).

[0071] like Figure 4 As shown, the next level of camouflage representation fragment rf n (n≥1, n represents the serial number corresponding to the camouflage target representation) and the camouflage identification feature r extracted at the previous level n+1 After combining, convolution is performed, conv represents convolution, and low-frequency component LL, high-frequency component LH, high-frequency component HL and high-frequency component HH are obtained;

[0072] The low-frequency component LL captures the camouflaged segmentation information and of the camouflaged target; the low-frequency component LL is enhanced by the attention mechanism AM and the instance normalized residual block ResIN, and the camouflaged segmentation information and contour information are superimposed to obtain the low-frequency component camouflage feature r LL ;

[0073] The high-frequency component LH, high-frequency component HL and high-frequency component HH are enhanced by the normalized residual block ResB and the attention mechanism AM respectively to obtain the high-frequency component camouflage feature r LH , high frequency component camouflage feature r HL and high frequency component camouflage feature r HH ;

[0074] The low frequency component disguises the feature r LL , high frequency component camouflage feature r LH , high frequency component camouflage feature r HLand high frequency component camouflage feature r HH Perform transposed convolution TC to obtain the camouflage recognition feature r n .

[0075] Convolving each camouflage recognition feature obtains a camouflage recognition level image, and superimposing each camouflage recognition level image obtains a camouflage recognition image.

[0076] The training process of the target search model and the target recognition model includes:

[0077] Collect training images containing camouflaged targets to obtain a training data set;

[0078] The target search model and target recognition model will be trained through the training data set, and the structured loss value S of camouflage recognition will be calculated during the training process. The parameters of the target search model and target recognition model will be iteratively adjusted according to the structured loss value S to obtain the target search model and target recognition model that meet the set requirements.

[0079] Calculate the structural loss value S of camouflage recognition during training. The calculation formula is:

[0080] S=α*S o +(1-α)*S r

[0081]

[0082] S o =μ*O FG +(1-μ)*O BG

[0083] Among them, S o It is expressed as an object-oriented structural similarity measure; S r It is represented as a region-oriented structural similarity measure; α is represented as a weight factor; W k represents the weight of the kth camouflage representation slice; K represents the total number of camouflage representation slices; k represents the sequence number of the camouflage representation slice; μ is the proportion of the foreground area in the image; O BG Denoted as background measure; O FG Denoted as a foreground metric.

[0084] Calculate the background metric O BG and the foreground metric O FG , the expression formula is:

[0085]

[0086]

[0087]

[0088]

[0089] Among them, D FG It is expressed as the difference between the foreground area of ​​the camouflage prediction map and the camouflage representation fragment, x FG Represents the probability value of the foreground area of ​​the camouflaged target prediction map, y FG They represent the probability values ​​of the foreground area of ​​the camouflage representation fragment, Expressed as probability values ​​x FG The mean of Expressed as probability values ​​y FG The mean of Expressed as probability values ​​x FG The variance, D BG It is expressed as the difference between the background area of ​​the camouflage prediction image and the camouflage representation fragment, x BG Represents the probability value of the background area of ​​the camouflaged target prediction map, y BG They represent the probability values ​​of the background area of ​​the camouflage representation fragment respectively, Expressed as probability values ​​x BG The mean of Expressed as probability values ​​y BG The mean of Expressed as probability values ​​x BG The variance of .

[0090] As attached Figure 3 As shown, an intuitive visual comparison of five mainstream target segmentation network models, SINet (Search Identification Network), Parallel Reverse Attention Network (PraNet), CPD (Cascaded Partial Decoder), PFANet (Pyramidal Feature Aggregation Network), and EGNet (Edge Guidance Network), wherein GT (Ground Truth) represents the segmentation truth value and Ours represents the camouflaged target detection method of the present invention; through the comparison, it can be seen that the present invention can restore a clearer camouflaged target image, and at the same time has better edge texture information, and can better detect it from an environment that is highly similar to the camouflaged target.

[0091] Embodiment 2

[0092] The second aspect of the present invention provides a disguised target detection system based on a learning neural network. The disguised target detection system provided in this embodiment can be applied to the disguised target detection method described in the first embodiment, including:

[0093] An acquisition module, which acquires detection images;

[0094] A target search module is used to input the collected detection image into a pre-trained target search model to obtain a multi-layer camouflage target representation arranged in sequence from low level to high level;

[0095] The target recognition module extracts camouflage representation fragments from each layer of camouflage target representation and inputs them into a pre-trained target recognition model to obtain multiple camouflage recognition features of corresponding levels;

[0096] A camouflage recognition image generation module convolves each camouflage recognition feature to obtain a camouflage recognition level image, and superimposes each camouflage recognition level image to obtain a camouflage recognition image;

[0097] The training module is used to train the target search model and the target recognition model.

[0098] The target search model includes setting a feature extraction layer Stage1, a feature extraction layer Stage2, a feature extraction layer Stage3, a feature extraction layer Stage4 and a feature extraction layer Stage5 in sequence; the feature extraction layer Stage2, the feature extraction layer Stage3, the feature extraction layer Stage4 and the feature extraction layer Stage5 extract camouflage representation f from the collected image in sequence; the feature extraction layer Stage2, the feature extraction layer Stage3, the feature extraction layer Stage4 and the feature extraction layer Stage5 are respectively connected to a receptive field network for amplifying local detail information of the camouflage representation f; the target recognition model includes four wavelet attention units for wavelet attention transformation; the four wavelet attention units are connected to four receptive field networks in a one-to-one correspondence;

[0099] The disguised representation fragment fr′5, the disguised representation fragment fr′4, the disguised representation fragment fr′3 and the disguised representation fragment fr′2 are sequentially input into four wavelet attention units in the order from low level to high level;

[0100] The wavelet attention unit performs wavelet attention transformation on the lowest level camouflage representation fragment fr′5 to extract the camouflage recognition feature r5;

[0101] Slice the next level of camouflage representation rf n and the camouflage identification feature r extracted at the previous level n+1 After the combination, the camouflage recognition features are obtained in sequence by performing wavelet attention transformation through the wavelet attention unit.

[0102] Embodiment 3

[0103] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the camouflaged target detection method described in the first embodiment.

[0104] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0105] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0106] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0107] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0108] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A camouflaged target detection method based on learning neural network, characterized in that: include: Collect the detection image and input it into the pre-trained target search model to obtain a multi-layer camouflage target representation that is set in sequence from low level to high level; Extract camouflage representation fragments from each layer of camouflage target representation and input them into a pre-trained target recognition model to obtain multiple camouflage recognition features of corresponding levels; Convolving each camouflage recognition feature to obtain a camouflage recognition level image, and superimposing each camouflage recognition level image to obtain a camouflage recognition image; The training process of the target search model and the target recognition model includes: Collect training images containing camouflaged targets to obtain a training data set; The target search model and target recognition model are trained through the training data set, and the structural loss value S of camouflage recognition during the training process is calculated. The calculation formula is: ； ； ； in, Expressed as an object-oriented structural similarity measure; It is expressed as a region-oriented structural similarity measure; Expressed as a weight factor; It represents the weight of the kth disguised representation fragment; K represents the total number of disguised representation fragments; k represents the sequence number of the disguised representation fragment; is the ratio of the foreground area to the image; It is represented as background metric; denoted as a prospect measure; Calculate the background metric and prospect metrics , the expression formula is: ； ； ； ； in, It is expressed as the difference between the foreground area of ​​the camouflage prediction map and the camouflage representation fragment. Represents the probability value of the foreground area of ​​the camouflaged target prediction map, They represent the probability values ​​of the foreground area of ​​the camouflage representation fragment, Expressed as probability values The mean of Expressed as probability values The mean of Expressed as probability values The variance of It is expressed as the difference between the background area of ​​the camouflage prediction image and the camouflage representation fragment. Represents the probability value of the background area of ​​the camouflaged target prediction map, They represent the probability values ​​of the background area of ​​the camouflage representation fragment respectively, Expressed as probability values The mean of Expressed as probability values The mean of Expressed as probability values The variance of The parameters of the target search model and the target recognition model are iteratively adjusted according to the structured loss value S to obtain the target search model and the target recognition model that meet the set requirements.

2. The method for detecting camouflaged targets based on learning neural networks according to claim 1, characterized in that: The method of collecting and detecting images and inputting them into a pre-trained target search model to obtain a multi-layer camouflage target representation arranged in sequence from low level to high level includes: The detection image is input into the Res2Net50 residual network in the target search model, and the four-layer camouflage representation f is output in order from low level to high level; The local detail information of each layer of camouflage representation f is amplified through the receptive field network to obtain camouflage target representation fr2, camouflage target representation fr3, camouflage target representation fr4 and camouflage target representation fr5 respectively.

3. The method for detecting camouflaged targets based on learning neural networks according to claim 2, characterized in that: The method of extracting camouflage representation fragments from each layer of camouflage target representation includes: Extract the four slices of the disguised target representation fr5 as the disguised representation slice ; Extract the two slices of the disguised target representation fr4 as the disguised representation slice ; Extract the entire camouflaged target representation fr3 as the camouflaged representation fragment ; Extract the entire camouflage target representation fr2 as the camouflage representation fragment .

4. The method for detecting camouflaged targets based on learning neural networks according to claim 3 is characterized in that: The method of inputting the camouflage representation fragments into a pre-trained target recognition model to obtain multiple camouflage recognition features of corresponding levels includes: Segment the camouflage representation in order from low to high , camouflage representation fragmentation , camouflage representation fragmentation and disguise representation fragmentation Input into the target recognition model in sequence; Segmentation of the lowest level of disguise representation Perform wavelet attention transform to extract camouflage recognition feature r5; Slice the next level of camouflage representation rf n and the camouflage identification feature r extracted at the previous level n+1 After combining, the camouflage recognition features r are obtained in turn through wavelet attention transform n .

5. The method for detecting camouflaged targets based on learning neural networks according to claim 4, characterized in that: Slice the next level of camouflage representation rf n and the camouflage identification feature r extracted at the previous level n+1 After combining, the methods of obtaining camouflage identification features in sequence through wavelet attention transform include: Slice the next level of camouflage representation rf n and the camouflage identification feature r extracted at the previous level n+1 After combining, convolution is performed to obtain low-frequency component LL, high-frequency component LH, high-frequency component HL and high-frequency component HH; The low-frequency component LL captures the camouflaged segmentation information and contour information of the camouflaged target; the low-frequency component LL is enhanced by the attention mechanism AM and the instance normalized residual block ResIN, and the camouflaged segmentation information and contour information are superimposed to obtain the low-frequency component camouflage feature r LL ; The high-frequency component LH, high-frequency component HL and high-frequency component HH are enhanced by the normalized residual block ResB and the attention mechanism AM respectively to obtain the high-frequency component camouflage feature r LH , high frequency component camouflage feature r HL and high frequency component camouflage feature r HH ; The low frequency component disguises the feature r LL , high frequency component camouflage feature r LH , high frequency component camouflage feature r HL and high frequency component camouflage feature r HH Perform transposed convolution TC to obtain the camouflage recognition feature r n .

6. A camouflaged target detection system based on learning neural network, characterized in that: include: An acquisition module, which acquires detection images; A target search module is used to input the collected detection image into a pre-trained target search model to obtain a multi-layer camouflage target representation arranged in sequence from low level to high level; The target recognition module extracts camouflage representation fragments from each layer of camouflage target representation and inputs them into a pre-trained target recognition model to obtain multiple camouflage recognition features of corresponding levels; A camouflage recognition image generation module convolves each camouflage recognition feature to obtain a camouflage recognition level image, and superimposes each camouflage recognition level image to obtain a camouflage recognition image; A training module, used for training the target search model and the target recognition model; The process of training the target search model and the target recognition model by the training module includes: Collect training images containing camouflaged targets to obtain a training data set; The target search model and target recognition model are trained through the training data set, and the structural loss value S of camouflage recognition during the training process is calculated. The calculation formula is: ； ； ； in, Expressed as an object-oriented structural similarity measure; It is expressed as a region-oriented structural similarity measure; Expressed as a weight factor; It represents the weight of the kth disguised representation fragment; K represents the total number of disguised representation fragments; k represents the sequence number of the disguised representation fragment; is the ratio of the foreground area to the image; It is represented as background metric; denoted as a prospect measure; Calculate the background metric and prospect metrics , the expression formula is: ； ； ； ； in, It is expressed as the difference between the foreground area of ​​the camouflage prediction map and the camouflage representation fragment. Represents the probability value of the foreground area of ​​the camouflaged target prediction map, They represent the probability values ​​of the foreground area of ​​the camouflage representation fragment, Expressed as probability values The mean of Expressed as probability values The mean of Expressed as probability values The variance of It is expressed as the difference between the background area of ​​the camouflage prediction image and the camouflage representation fragment. Represents the probability value of the background area of ​​the camouflaged target prediction map, They represent the probability values ​​of the background area of ​​the camouflage representation fragment respectively, Expressed as probability values The mean of Expressed as probability values The mean of Expressed as probability values The variance of The parameters of the target search model and the target recognition model are iteratively adjusted according to the structured loss value S to obtain the target search model and the target recognition model that meet the set requirements.

7. A camouflaged target detection system based on learning neural network according to claim 6, characterized in that: The target search model includes setting a feature extraction layer Stage1, a feature extraction layer Stage2, a feature extraction layer Stage3, a feature extraction layer Stage4 and a feature extraction layer Stage5 in sequence; the feature extraction layer Stage2, the feature extraction layer Stage3, the feature extraction layer Stage4 and the feature extraction layer Stage5 extract camouflage representation f from the collected image in sequence; the feature extraction layer Stage2, the feature extraction layer Stage3, the feature extraction layer Stage4 and the feature extraction layer Stage5 are respectively connected to a receptive field network for amplifying local detail information of the camouflage representation f; The target recognition model includes four wavelet attention units for wavelet attention transformation; the four wavelet attention units are connected to four receptive field networks in a one-to-one correspondence.

8. A computer-readable storage medium, characterized in that A computer program is stored thereon, and when the program is executed by a processor, the steps of the camouflaged target detection method described in any one of claims 1 to 5 are implemented.

Images