A camouflaged target detection method and system simulating biological perception

Abstract

The present invention discloses a camouflaged target detection method and system that simulates biological perception. According to the process of biological perception of camouflaged targets, neural network modeling is performed. The modeling includes a five-stage camouflaged target recognition process: information collection, information enhancement, information filtering, information positioning, information correction and prediction. Customized information processing mechanisms of feature extraction network, information enhancement component, information filtering component, information positioning component and information correction component are respectively configured. Then, the loss function used in model training is constructed, the hyperparameters of model training, the enhancement method and size of the training image are determined; finally, the model is trained, and the trained optimal model is used to detect the camouflaged target image and check the effect, so as to accurately identify the camouflaged target.

Description

Technical Field

[0001] The present invention belongs to the technical field of target detection, and mainly relates to a camouflaged target detection method and system that simulates biological perception. Background Art

[0002] Camouflaged object detection is an extension of salient object detection and general object detection. Unlike these two approaches, it focuses on identifying objects that blend in with their surroundings. Common camouflaged creatures, such as chameleons and insects, blend in with their surroundings by changing physical features like color or shape to deceive predators. There are also artificially camouflaged objects, such as camouflage uniforms and body paint, which blend the human body into the environment. Camouflaged object detection is a benchmark problem in object detection and has significant practical value in scenarios such as military applications, bioconservation, agricultural applications, and medical diagnosis.

[0003] In recent years, methods based on deep learning have been widely used in the field of computer vision and are considered the preferred approach for detecting camouflaged targets. Since camouflaged targets are often hidden in the background, this makes identification more challenging. Therefore, many approaches attempt to mimic the behavior of humans and animals in detecting camouflage by constructing convolutional neural networks. For example, some approaches model the recognition process as a two-stage process: search followed by recognition. Others apply geometric transformations to images, such as rotation, reduction, and magnification, to model the processing mechanisms of perceptual information. However, these modeling approaches still fail to fully model the behavior of humans and animals in detecting camouflaged targets, nor do they reconstruct the close relationship between perceptual behavior and perceptual information processing mechanisms. Therefore, how to enable these modeled neural networks to effectively perceive camouflaged targets, similar to biological processes, is a question worthy of further consideration. Summary of the Invention

[0004] The present invention is aimed at the problem that the existing technology lacks a point-to-point relationship between perception behavior and perception information processing. It provides a camouflaged target detection method and system that simulates biological perception. According to the process of biological perception of camouflaged targets, neural network modeling is performed. The modeling includes a five-stage camouflaged target recognition process: information collection, information enhancement, information filtering, information positioning, information correction and prediction, and respectively configures customized information processing mechanisms of feature extraction network, information enhancement component, information filtering component, information positioning component and information correction component, then constructs the loss function used in model training, determines the hyperparameters of model training, and the enhancement method and size of the training image; finally, the model is trained, and the trained optimal model is used to detect the camouflaged target image and view the effect, so as to accurately identify the camouflaged target.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is: a camouflaged target detection method simulating biological perception, comprising the following steps:

[0006] S1: Based on the process of biological perception of camouflaged targets, a neural network model is built. The model includes a five-stage camouflaged target recognition process: information collection, information enhancement, information filtering, information positioning, information correction, and prediction;

[0007] S2: Configure customized information processing mechanisms for each of the five stages in step S1, namely: feature extraction network, information enhancement component, information filtering component, information positioning component, and information correction component;

[0008] S3: Construct a loss function for model training. The loss function uses the true label as supervision information, including weighted binary cross entropy loss and weighted intersection-over-union loss.

[0009] S4: Determine the hyperparameters for model training and the enhancement method and size of the training images;

[0010] S5: Train the model and use the trained optimal model to detect the disguised target image and check the effect.

[0011] As an improvement of the present invention, in the information collection stage, a pre-trained feature extraction network is used to simulate biological scanning images to capture image information, and the feature extraction network is a convolutional neural network or a Transformer-type network.

[0012] As an improvement of the present invention, in the information enhancement stage, an information enhancement component is constructed to simulate the biological observation of as much information as possible from different distances. The information enhancement component uses dilated convolution, ordinary convolution, and pooling operations to construct a five-branch multi-receptive field information capture mechanism, and the features pre-extracted in the information collection stage are reviewed under multiple receptive fields.

[0013] As another improvement of the present invention, in the information filtering stage, an information filtering component is constructed to model the organism to filter out irrelevant information. The information filtering component groups features through shuffle and split, and then completes the filtering through average group pooling and maximum group pooling.

[0014] As another improvement of the present invention, in the information localization stage, an information localization component is constructed to model the biological target information from both global and local perspectives. The information localization component includes a dual-branch structure, one branch is responsible for capturing the global information of the target, and the global information extraction branch is composed of Multi-Dconv Head Transposed Attention; the other branch is responsible for capturing the local information of the target.

[0015] As another improvement of the present invention, in the information correction and prediction stage, an information correction component is constructed to model the organism to adjust the local prediction information multiple times. The information correction component divides the positioning information into two parts, certain and uncertain, through uncertain area calculation. Channel attention is used to assign different weights to different certain information in the channel dimension, and spatial attention is used to assign different weights to different uncertain information in the spatial dimension. Finally, the corrected information is obtained by fusing it with the input features that have undergone both spatial and channel attention.

[0016] As a further improvement of the present invention, in step S3, the weighted binary cross entropy loss The specific calculation method is as follows:

[0017]

[0018] Where H and W represent the height and width of the image, respectively, ρ(·) is the indicator function, γ is the hyperparameter coefficient, l∈{0,1} is the label of the two categories, and Represents the predicted value and true value at position (i, j), α ij represents the importance of the pixel at position (i, j), θ represents all model parameters, and Pr(·) represents the predicted probability;

[0019] The weighted intersection-over-union loss Specifically:

[0020]

[0021] Therefore, the final training loss is calculated as follows:

[0022]

[0023] Where, P i and G are the predicted result graph and the labeled true graph of each branch respectively.

[0024] As a further improvement of the present invention, in step S4, the hyperparameters of the model training are specifically: the learning rate is 1e-4, the total number of training rounds is 150 epochs, and the learning rate is reduced by 10 times at the 30th, 40th, and 50th epochs; the size of the training image is specifically a mixed training of three resolutions: 384×384, 512×512, and 704×704.

[0025] In order to achieve the above-mentioned purpose, the present invention also adopts a technical solution: a camouflaged target detection system that simulates biological perception, including a computer program, which implements the steps of any of the above-mentioned methods when executed by a processor.

[0026] Compared with the existing technology, the present invention has the following beneficial effects: the present invention discloses a method and system for detecting camouflaged targets that simulates biological perception, and designs a more comprehensive and detailed neural network modeling method. This method divides the behavior of biological perception of camouflaged targets into five stages, namely information collection, information enhancement, information filtering, information positioning, and information correction and prediction. Then, a corresponding perception information processing mechanism is customized for each stage, namely, an information enhancement component, an information filtering component, an information positioning component, and an information correction component. In addition, the method can flexibly replace the perception information processing mechanism of each stage according to the performance of each component. Through this flexible and organic combination of perception behavior and perception information processing mechanism, the method more vividly reproduces the entire process of biological perception of camouflaged targets and significantly improves the performance of camouflaged target detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is a schematic flow chart of the steps of a method for detecting camouflaged targets simulating biological perception according to the present invention;

[0028] Figure 2 This is a schematic diagram of the structure of an information enhancement component according to an embodiment of the present invention;

[0029] Figure 3 This is a schematic diagram of the structure of an information filtering component according to an embodiment of the present invention;

[0030] Figure 4 This is a schematic diagram of the structure of an information positioning component according to an embodiment of the present invention;

[0031] Figure 5 Schematic diagram of the structure of the information correction component of an embodiment of the present invention. DETAILED DESCRIPTION

[0032] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0033] Example 1

[0034] A method for detecting camouflaged targets by simulating biological perception comprises the following steps:

[0035] Step S1: Analyze the behavior of biological perception of disguised targets and model a five-stage disguised target recognition process, which is: information collection, information enhancement, information filtering, information positioning, information correction and prediction;

[0036] like Figure 1 As shown, Figure 1 The one-to-one correspondence process of the method of the present invention in simulating the biological perception process is shown;

[0037] The five-stage one-to-one correspondence modeling process is specifically manifested as follows: (1) information acquisition stage, in which high-dimensional features of images are extracted by pre-trained convolutional neural networks or Transformer-type networks to model the behavior of organisms reading image information; (2) information enhancement stage, in which the extracted features are enhanced by the proposed information enhancement component to obtain sufficient and information-rich features to model the behavior of organisms transforming observations; (3) information filtering stage, in which the extracted rich features are screened by the proposed information filtering component to model the behavior of organisms processing useful and useless information; (4) information The localization stage uses the proposed global-local component to focus on the filtered information from both global and local perspectives, thereby modeling the biological behavior of processing key information globally and locally and discriminating the final target location information; (5) Information correction and target prediction: In this stage, an information correction component is first constructed to evaluate the localized target information and perform local corrections to simulate the behavior of the biological multiple adjustments to the identified target location information; then a convolution is used to summarize the corrected target information and reduce the dimension to output the target prediction result, thereby modeling the behavior of the biological use of high-order perceptual information to finally capture the camouflaged target. Each of these five stages has its own targeted perceptual information processing mechanism, which effectively solves the problem of separation or weak coupling between perceptual behavior and perceptual information processing.

[0038] Step S2: Based on the five-stage modeling scheme designed to simulate biological perception of camouflaged targets, this step designs customized information processing mechanisms for each stage of the five-stage perception process, namely: basic feature extraction network, information enhancement component, information filtering component, information positioning component, and information correction component;

[0039] During the information collection phase, a pre-trained feature extraction network is used to simulate biological scanning images to capture image information. The basic feature extraction network can be a pre-trained network of the ResNet or Transformer class. In this embodiment, the Transformer variant PVT-V2 is selected as the feature extractor to capture the feature information of the image.

[0040] During the information enhancement phase, an information enhancement component is constructed to simulate an organism's ability to observe as much information as possible from different distances. When searching for a target, an organism will move closer or further away from the observed target and the surrounding environment to obtain information from multiple perspectives to help find the target. The information enhancement component reviews the pre-extracted features under multiple receptive fields, effectively modeling this behavior. The information enhancement component considers capturing features from multiple receptive fields. Specifically, it uses a five-branch multi-receptive field information capture mechanism constructed using dilated convolution, ordinary convolution, and pooling operations. In addition, considering the hole filling of dilated convolution, an additional three-layer stacked CBR layer is designed to eliminate grid effects.

[0041] For the information enhancement component in this embodiment, please refer to Figure 2 Structural diagram of, given input feature Input∈R C×H×W , the first branch uses CBR with a convolution kernel of 1×1 (composed of a convolution layer, a BatchNormalize layer, and a ReLU layer) to extract the feature f1 under a very small field of view, and the last branch uses an adaptive average pooling layer and a CBR layer with a convolution kernel of 1×1 to extract the feature f5 of the global field of view. The other three branches are composed of three stacked CBR layers. The first branch extracts the feature f2 by a 3-layer CBR with a convolution kernel size of 3×3. In order to extract features from a changing receptive field and avoid grid effects, the last two branches use a combination of dilated convolutions with dilation rates of [1, 2, 3] and [2, 4, 16] to extract features f3 and f4 with convolution kernel sizes of 3×3 and 5×5, respectively. In addition, we capture the input and the features f extracted by each branch i The information residual between (i∈[1,5]) is used as additional enhanced supplementary information (denoted as res i i∈[1,5]). The final output information enhanced by the information enhancement component is Output∈R C×H×W It can be expressed as:

[0042] Output=F conv1 (F cat (f1,f2,f3,f4,f5))+F conv1 (F cat (res1,res2,res3,res4,res5))

[0043] Among them F conv1 and F cat They represent the connection operation and the convolution operation with a convolution kernel of 1×1, respectively.

[0044] During the information filtering phase, an information filtering component is constructed to model how organisms filter out irrelevant information. After acquiring multiple pieces of information, organisms process it to retain key information and remove irrelevant information to help accurately locate targets. The information filtering component filters the rich features extracted from various receptive fields to obtain representative information to model this perception behavior. The core of the information filtering component is average group pooling and maximum group pooling, which are used to select the most representative features to better identify camouflaged targets. In addition, shuffle and split are used before filtering to better group features.

[0045] For the information filtering component in this embodiment, please refer to Figure 3 The information filtering component is divided into two main stages: 1. Feature preparation: Input feature Input∈R C×H×WFirst, the number of channels is further expanded by n times through a 1×1 convolution layer to obtain f expand In preparation for screening, the shuffle operation is then used in the channel dimension to expand Shuffle to increase the diversity of features within the group after grouping, and finally divide the shuffled features into m group features f using group-wise operations i (i∈{1,2…,m}). The formula can be expressed as:

[0046]

[0047] Among them F group and F shuffle denote the grouping and channel shuffling operations respectively, and ∪(·) denotes the integration operation.

[0048] 2. Feature filtering: based on f i , we designed two group-wise pooling layers, namely channel maximum pooling (CMaxpooling) and channel average pooling (CMeanpooling), to obtain the maximum response of the features within the group and the average response Followed by a concatenate operation and a convolution layer with a convolution kernel of 3×3 and The feature Output∈R after fusion output filtering C×H×W ,The process can be expressed as:

[0049] Output=F conv3 (F cat (f max ,f mean ))+Input

[0050] in

[0051]

[0052]

[0053] Where M(·) represents the maximum value operation.

[0054] During the information localization phase, an information localization component is constructed to model how organisms locate target information globally and locally. To locate the target, organisms focus on key information and conduct global and local observations to obtain the target's overall and detailed information, and then accurately identify the target based on the localization information. The information localization component processes the filtered information from two branches, namely global localization and local detail search, to accurately locate the target. Specifically, the information localization component has a dual-branch structure. One branch is responsible for capturing the global information of the target. The global information extraction branch is composed of Multi-Dconv Head Transposed Attention. The other branch is responsible for capturing the local information of the target, mainly through feature partitioning to increase feature diversity while reducing the number of parameters. Ultimately, this dual-branch structure supports the model in recognizing the overall and detailed information of the target.

[0055] For the information location component in this embodiment, please refer to Figure 4 The left branch uses Multi-Dconv HeadTransposed Attention from Input∈R C×H×W Extract global position information f g To locate the target position, the right branch extracts local detail information f through a local module l To refine the location of the target. Specifically, the local detail extraction module consists of three branches. The first two branches use a split operation to divide the input feature input into two parts by channel. Two parallel CBRs are then used to extract different local information. These two parts of information are then fused using a concatenation layer with a kernel size of 1×1. This reduces the number of parameters and increases feature diversity.

[0056] In particular, in order to capture detailed information of different sizes, the two branches use convolution kernels of two different sizes: 3×3 and 5×5, respectively.

[0057] In addition, the feature division operation hinders the interaction between the two parts of the features, so a third branch is added to compensate for the feature interaction information between all channels. The local detail information is the pixel-by-pixel sum of the information f1, f2, and f3 of the three branches, and finally captures the positioning information of the target Output∈R C×H×W It can be expressed as:

[0058] Output=F conv3 (f g +f l )+Input.

[0059] During the information correction and prediction phase, an information correction component is constructed to model the biological process of repeatedly adjusting local prediction information. Biological processes perform multiple local corrections on positioning information to improve positioning accuracy. Based on this behavior, the information correction component calculates the previously uncertain positioning information to perform local corrections. Specifically, the information correction component divides positioning information into two parts: certain and uncertain. It uses channel attention to assign different weights to different certain information in the channel dimension, and spatial attention to assign different weights to different uncertain information in the spatial dimension. Finally, the corrected information is fused with the input features that have undergone both spatial and channel attention.

[0060] For the information correction component of this embodiment, please refer to Figure 5 Structure diagram, input feature Input∈R C×H×W The information is then fed into the Residual Spatial Attention Module (RSAB) and the Residual Channel Attention Module (RCAB) to perform varying degrees of attention on the current overall positioning information from both spatial and channel dimensions, thereby determining the importance of each piece of information to the predicted target. Furthermore, the information correction component uses the previous perception results to determine the areas where correction may or may not be required for the current positioning information. Specifically, through the Uncertain Region Computation (URC), the previous prediction result mask is divided into an Uncertain Region (UR) and a Certain Region (1-UR). The calculation process is as follows:

[0061]

[0062] Then, the current positioning information is divided into certain positioning information and uncertain positioning information through element-wise multiplication. The uncertainty of the uncertain positioning information is mainly reflected in the space, so RSAB is used to determine the importance of each uncertain information in the spatial dimension. The certain positioning information is certain in space, so RCAB is used to determine the importance of each certain information in the channel dimension. The final corrected information can be expressed as:

[0063] Output=F cat (f ur ,f ir ,F RSAB (Input)+F RCAB (Input)

[0064] in,

[0065]

[0066]

[0067] Finally, the output corrected positioning information is passed through a single-layer convolution layer with a kernel size of 3×3 and an upsampling operation to obtain the current prediction result.

[0068] Step S3: construct a loss function for model training;

[0069] The loss function used for detection uses the true label as supervision information, which includes: weighted binary cross entropy loss Weighted Intersection-over-Union Loss

[0070] This implementation uses weighted binary cross entropy loss Optimize the local features of the prediction results, and the calculation formula is as follows:

[0071]

[0072] Where H and W represent the height and width of the image, respectively, ρ(·) is the indicator function, γ is the hyperparameter coefficient, l∈{0,1} is the label of the two categories, and Represents the predicted value and true value at position (i, j), α ij represents the importance of the pixel at position (i, j), θ represents all model parameters, and Pr(·) represents the predicted probability.

[0073] This implementation additionally uses weighted intersection-over-union loss Considering the global characteristics to optimize the prediction results, the calculation formula is as follows:

[0074]

[0075] Therefore, the final training loss calculation formula is:

[0076]

[0077] Where, P i and G are the predicted result graph and the labeled true graph of each branch respectively.

[0078] Step S4: determine the hyperparameters for model training and the enhancement method and size of the training image;

[0079] This implementation determined the model training hyperparameters through multiple comparative experiments. Specifically, the learning rate was 1e-4, the total training round was 150 epochs, and the learning rate was reduced by a factor of 10 at the 30th, 40th, and 50th epochs. The optimizer used was the AdamW optimizer, which is commonly used in Transformer networks.

[0080] Furthermore, during experiments, this embodiment found that using high-resolution images for training can improve model performance, but it can also sometimes reduce the performance of some low-resolution test sets. Therefore, this embodiment employs a mixed training scheme with three resolutions: 384×384, 512×512, and 704×704, taking into account data of different resolutions and significantly improving model performance. Furthermore, based on practical experience, this embodiment employs two data augmentation methods: horizontal flipping and random rotation to increase the diversity of the training set.

[0081] Step S5: train the model, and use the trained optimal model to detect the disguised target image and check the effect.

[0082] In this implementation, all parameters were adjusted to their optimal values ​​and the model was retrained to obtain the optimal model. Finally, the model performance was evaluated through experiments and analysis. The final results showed that the proposed method for camouflaged target detection simulating biological perception significantly improved the accuracy of camouflaged target detection.

[0083] In summary, the present invention discloses a camouflaged target detection method that simulates biological perception, which divides the entire perception process into five stages: information collection, information enhancement, information filtering, information positioning, information correction and prediction, and designs a customized perception information processing mechanism for each stage, which are the basic feature extraction network, information enhancement component, information filtering component, information positioning component, and information correction component, to construct a point-to-point relationship between perception behavior and perception information processing, to model the tight coupling relationship between the two, effectively reproduce the behavior of organisms discovering camouflaged targets, and significantly improve the accuracy of camouflaged target detection.

[0084] It should be noted that the above content merely illustrates the technical idea of ​​the present invention and cannot be used to limit the scope of protection of the present invention. For ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications all fall within the scope of protection of the claims of the present invention.

Claims

1. A method for detecting camouflaged targets by simulating biological perception, characterized in that: The process includes the following steps: S1: Based on the process of biological perception of camouflaged targets, a neural network model is built. The model includes a five-stage camouflaged target recognition process: information collection, information enhancement, information filtering, information positioning, information correction, and prediction; S2: Configure customized information processing mechanisms for each of the five stages in step S1, namely: feature extraction network, information enhancement component, information filtering component, information positioning component, and information correction component; S3: Construct a loss function for model training. The loss function uses the true label as supervision information, including weighted binary cross entropy loss and weighted intersection-over-union loss. S4: Determine the hyperparameters for model training and the enhancement method and size of the training images; S5: Train the model and use the trained optimal model to detect the disguised target image and check the effect.

2. The method for detecting a camouflaged target by simulating biological perception according to claim 1, wherein: In the information collection stage, a pre-trained feature extraction network is used to simulate biological scanning images to capture image information. The feature extraction network is a convolutional neural network or a Transformer-type network.

3. The method for detecting a camouflaged target by simulating biological perception according to claim 1, wherein: In the information enhancement stage, an information enhancement component is constructed to simulate the biological observation of as much information as possible from different distances. The information enhancement component uses dilated convolution, ordinary convolution, and pooling operations to construct a five-branch multi-receptive field information capture mechanism, and re-examines the features pre-extracted in the information collection stage under multiple receptive fields.

4. The method for detecting a disguised target by simulating biological perception according to claim 1, wherein: In the information filtering stage, an information filtering component is constructed to model the organism to filter out irrelevant information. The information filtering component groups features through shuffle and split, and then completes the filtering through average group pooling and maximum group pooling.

5. The method for detecting a disguised target by simulating biological perception according to claim 1, wherein: In the information localization stage, an information localization component is constructed to model the biological global and local positioning target information. The information localization component includes a dual-branch structure, one branch is responsible for capturing the global information of the target, and the global information extraction branch is composed of Multi-DconvHead Transposed Attention; the other branch is responsible for capturing the local information of the target.

6. The method for detecting a disguised target by simulating biological perception according to claim 1, wherein: During the information correction and prediction stage, an information correction component is constructed to model the organism and adjust the local prediction information multiple times. The information correction component divides the positioning information into two parts, certain and uncertain, through uncertain area calculation. Channel attention is used to assign different weights to different certain information in the channel dimension, and spatial attention is used to assign different weights to different uncertain information in the spatial dimension. Finally, the corrected information is obtained by fusing it with the input features that have undergone both spatial and channel attention.

7. A method for detecting a disguised target by simulating biological perception according to any one of claims 2 to 6, characterized in that: In step S3, the weighted binary cross entropy loss The specific calculation method is as follows: Where H and W represent the height and width of the image, respectively, ρ(·) is the indicator function, γ is the hyperparameter coefficient, l∈{0,1} is the label of the two categories, and Represents the predicted value and true value at position (i, j), α ij represents the importance of the pixel at position (i, j), θ represents all model parameters, and Pr(·) represents the predicted probability; The weighted intersection-over-union loss Specifically: Therefore, the final training loss is calculated as follows: Where, P i and G are the predicted result graph and the labeled true graph of each branch respectively.

8. The method for detecting a disguised target by simulating biological perception according to claim 7, wherein: In step S4, the hyperparameters of the model training are as follows: a learning rate of 1e-4, a total number of training rounds of 150 epochs, and a learning rate reduction of 10 times at the 30th, 40th, and 50th epochs; and a mixed training of three resolutions of 384×384, 512×512, and 704×704 is performed on the training images.

9. A camouflaged target detection system simulating biological perception, comprising a computer program, characterized in that: When the computer program is executed by a processor, the steps of any one of the above methods are implemented.

Images