Camouflage target detection method and device based on subject and boundary perception

Abstract

The present invention discloses a method and device for detecting camouflaged targets based on subject and boundary perception, comprising: step S1, obtaining an environment image to be detected; step S2, preprocessing the environment image to be detected; step S3, inputting the preprocessed environment image to be detected into a camouflaged detection model for camouflaged target detection; wherein the camouflaged detection model comprises: a subject-boundary exploration module RBE, a subject enhancement module REM and a boundary optimization module BRD. The technical solution of the present invention can accurately locate and identify camouflaged targets.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and in particular relates to a method and device for detecting a camouflaged target based on subject and boundary perception. Background Art

[0002] Camouflage is the behavior of an object using various strategies to hide itself or deceive the observer, making it difficult to be detected. Camouflage comes in various forms, usually including changes in color, texture, and shape. Its purpose is to enhance the similarity between the object itself and the surrounding environment, thereby reducing the probability of being detected. For example, chameleons and soldiers wearing camouflage uniforms are typical examples of camouflage. The purpose of camouflaged object detection is to separate the camouflaged object from the environment that has a high degree of similarity with it. Due to the high similarity between the camouflaged object and the surrounding environment, it is currently impossible to distinguish the camouflaged object in a complex background. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide a method and device for detecting a camouflaged target based on subject and boundary perception, which can accurately locate and identify the camouflaged target.

[0004] To achieve the above object, the present invention adopts the following technical solution:

[0005] A camouflaged target detection method based on subject and boundary perception, comprising:

[0006] Step S1, obtaining an environment image to be detected;

[0007] Step S2, preprocessing the environment image to be detected;

[0008] Step S3: input the preprocessed environment image to be detected into the camouflage detection model to detect the camouflaged target; wherein the camouflage detection model includes: a body-boundary exploration module RBE, a body enhancement module REM and a boundary optimization module BRD.

[0009] Preferably, step S2 comprises:

[0010] Step S21: select PVTv2-B2 as the feature extraction network, input the environment image I to be detected into the feature extraction network to obtain multi-scale features Where i represents the number of layers in the feature extraction network, H and W are the length and width of the feature, respectively, and C i represents the channel dimension of the features in the i-th layer of the feature extraction network, The dimension of the vector representing the feature is H / 2 i ×W / 2 i ×C i ;

[0011] Step S22: Input MRFB module and get

[0012] Preferably, in step S3, I i 'Enter RBE to explore subject information and boundary information Will I i 'Aggregate with the output of the previous REM layer and and the output of the previous BRD are input into REM, wherein the fourth-layer REM has no previous BRD and previous REM, and only I'4 and As input; after each layer of REM, a BRD is set to enhance the boundary information again, and the output of each BRD passes through a segmentation head to obtain the prediction map and Supervised by region binary truth map, It is supervised by the boundary binarization truth map.

[0013] As a preferred method, the total loss function of the camouflage detection model is for:

[0014]

[0015] in, is the binary truth map of boundary information, is the Dice loss, is the mixed loss function, It consists of weighted BCE loss and weighted IoU loss.

[0016]

[0017] Among them, P represents the prediction graph, Represents the region binarization truth map.

[0018] The present invention provides a camouflaged target detection method based on subject and boundary perception, comprising:

[0019] An acquisition module, used for acquiring an environment image to be detected;

[0020] A preprocessing module, used for preprocessing the environment image to be detected;

[0021] The detection module is used to input the preprocessed environment image to be detected into the camouflage detection model for camouflage target detection; wherein the camouflage detection model includes: a body-boundary exploration module RBE, a body enhancement module REM and a boundary optimization module BRD.

[0022] Preferably, the preprocessing module comprises:

[0023] The first extraction unit is used to use PVTv2-B2 as a feature extraction network, input the environment image I to be detected into the feature extraction network to obtain multi-scale features Where i represents the number of layers in the feature extraction network, H and W are the length and width of the feature, respectively, and C i represents the channel dimension of the features in the i-th layer of the feature extraction network, The dimension of the vector representing the feature is H / 2 i ×W / 2 i ×C i ;

[0024] The second extraction unit will Input MRFB module and get

[0025] The invention preprocesses the environment image to be detected; inputs the preprocessed environment image to be detected into a camouflage detection model to detect the camouflage target; wherein the camouflage detection model includes: a body-boundary exploration module RBE, a body enhancement module REM and a boundary optimization module BRD. The technical solution of the invention can accurately locate and identify the camouflage target and enhance the boundary detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0027] Figure 1 It is a flow chart of a camouflaged target detection method based on subject and boundary perception according to an embodiment of the present invention;

[0028] Figure 2 It is a schematic diagram of the structure of the ABENet model;

[0029] Figure 3 This is the RBE module block diagram;

[0030] Figure 4 It is the block diagram of REM module;

[0031] Figure 5 This is the block diagram of the BRD module. DETAILED DESCRIPTION

[0032] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0033] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0034] Embodiment 1:

[0035] like Figure 1 , 2 As shown, an embodiment of the present invention provides a camouflaged target detection method based on subject and boundary perception, including:

[0036] Step S1, obtaining an environment image to be detected;

[0037] Step S2, preprocessing the environment image to be detected;

[0038] Step S3: input the preprocessed environment image to be detected into the camouflage detection model ABENet for camouflage target detection; wherein the camouflage detection model ABENet includes: a body-boundary exploration module (RBE), a body enhancement module (REM) and a boundary optimization module (BRD).

[0039] Further, step S2 is specifically as follows:

[0040] Step S21: select PVTv2-B2 as the feature extraction network, input the environment image I to be detected into the feature extraction network to obtain multi-scale features Where i represents the number of layers in the feature extraction network, H and W are the length and width of the feature, respectively, and C i represents the channel dimension of the features in the i-th layer of the feature extraction network, The dimension of the vector representing the feature is H / 2 i ×W / 2 i ×C i ;

[0041] Step S22: In order to reduce the network parameters and computational complexity, Input MRFB module to reduce the channel dimension and get

[0042] Further, in step S3, I i 'Enter RBE to explore subject information and boundary information In order to locate and identify the camouflaged target and enhance the main part of the camouflaged target, I i 'Aggregate with the output of the previous REM layer and The output of the previous BRD is input into REM together. The fourth-layer REM does not have the previous BRD and the previous REM. Therefore, only I'4 and I'4 are set for the fourth-layer REM. As input, REM uses a multi-scale approach to better locate camouflaged targets of different sizes and The output of the previous BRD is introduced to enhance the boundary information for the first time. After each layer of REM, a BRD is set to enhance the boundary information again, and the output of each BRD will pass through a segmentation head to obtain the prediction map. Introducing deep supervision strategies to better optimize the model, and Supervised by region binary truth map, It is supervised by the boundary binarization truth map.

[0043] As an implementation method of an embodiment of the present invention, the specific processing process of the subject-boundary exploration module (RBE) is as follows:

[0044] For the problem of blurred boundaries, currently, multi-scale features are directly cascaded to generate boundary clues. However, multi-scale features have the problem of spatial misalignment, and directly cascading multi-scale features may introduce additional noise that affects model performance. Inspired by the fact that after the image is upsampled, the main information remains basically consistent, but may be blurred at the boundary. The embodiment of the present invention RBE is used to better extract the main information and boundary information.

[0045] like Figure 3 As shown, the high-level features I3' and I'4 are input into the MPD module to obtain the feature F with subject information corresponding to the high-level features ch , and then pass through a 3×3 convolution and Sigmoid function to generate the main information Subject information Multiply by I1' and I'2 respectively and introduce residual connection to enhance the main information of low-level features, and then input the enhanced features into the MPD module to obtain the feature F with main information corresponding to the low-level features cl . ch Upsample to F cl The same size is then subtracted to obtain the information difference between the two scale features at the boundary, and then the subtracted features are passed through a 3×3 convolution and a Sigmoid function to generate boundary information

[0046] F ch=MPD(I3',I'4)·(1)

[0047]

[0048]

[0049]

[0050] Among them, Conv3(·) represents 3×3 convolution, σ(·) represents Sigmoid function, MPD(·) represents the above MPD module, and |·| represents taking the absolute value. In this way, the main information and boundary information can be better obtained to guide the subsequent modules to enhance the features.

[0051] As an implementation method of the embodiment of the present invention, the specific processing process of the subject enhancement module (REM) is as follows:

[0052] In order to help the network locate and detect targets more effectively, REM is used to enhance the main information of the target and the boundary information for the first time. Figure 2 As shown, REM has three inputs in total, Output of the previous BRD and subject information in, The corresponding I i ' and the output of the previous REM Aggregate from:

[0053]

[0054] in, represents two feature cascades along the channel direction, Conv3(·) represents a 3×3 convolution, and CBR(·) represents a convolution layer, including 3×3 convolution, batch normalization, and ReLu function. Figure 4 It can be seen that REM consists of two branches. The branches are first processed using dilated convolutions with different dilation rates (1, 3, 5, 7). Then the results of the dilated convolutions with different expansion rates are cascaded along the channel and passed through a convolution layer to obtain In this way, camouflaged targets of different sizes can be located and identified. and Multiply, then By retaining the original information through residual connection, the branch result can be obtained for Branch, let it pass through two convolutional layers, and then get the result of the branch In order to preserve the original feature information, and Cascade along the channel direction, then pass through a convolution layer and a 3×3 convolution to obtain The above process is formulated as:

[0055]

[0056]

[0057]

[0058] For aggregation The global and local information of Through global attention and local attention operations respectively. Specifically, for local attention, After CBR(·) and CBS(·) to obtain the local attention map, Multiply by the local attention map and add The characteristics of local information can be obtained As for global attention, unlike local attention, it undergoes a global average pooling operation before CBR(·) and CBS(·) to obtain the features of global information. Then and Adding and passing through a convolutional layer can get the output of the REM module

[0059]

[0060]

[0061]

[0062] Among them, CBS(·) includes a 3×3 convolution, batch normalization and Sigmoid function, and GAP(·) represents the global average pooling operation. For the 4th layer REM, since there is no output of the previous level BRD, The branch where There is no output from the previous REM, which is equal to I'4.

[0063] As an implementation method of the present invention, the specific processing process of the boundary optimization module (BRD) is as follows: Figure 5 As shown:

[0064] A BRD is set after REM to enhance the boundary information and decoding again. Specifically, the output of REM is first With boundary information Multiply and add To generate Next, Spatial attention is input to reduce background noise interference. Then, The result of spatial attention is cascaded along the channel and then passes through a convolution layer and 3×3 convolution to obtain the output result of BRD.

[0065]

[0066]

[0067] Among them, SAM(·) represents spatial attention.

[0068] As an implementation method of an embodiment of the present invention, the loss function of ABENet is as follows:

[0069] ABENet includes two kinds of supervision: regional information supervision and boundary information supervision.

[0070] For regional information, a hybrid loss function is used to assign more weight to harder pixels. It consists of weighted BCE loss and weighted IoU loss:

[0071]

[0072] Among them, P represents the prediction graph, Represents the region binarization truth map.

[0073] For regional information, select the prediction map obtained by a segmentation head from the output of each BRD And regional information Carry out supervision.

[0074] For boundary information, choose Dice loss To the boundary information Therefore, the total loss function of ABENet is The definition is as follows:

[0075]

[0076] in, It is the binary truth map of boundary information.

[0077] Embodiment 2:

[0078] The embodiment of the present invention further provides a camouflaged target detection method based on subject and boundary perception, comprising:

[0079] An acquisition module, used for acquiring an environment image to be detected;

[0080] A preprocessing module, used for preprocessing the environment image to be detected;

[0081] The detection module is used to input the preprocessed environment image to be detected into the camouflage detection model for camouflage target detection; wherein the camouflage detection model includes: a body-boundary exploration module RBE, a body enhancement module REM and a boundary optimization module BRD.

[0082] As an implementation of an embodiment of the present invention, the preprocessing module includes:

[0083] The first extraction unit is used to use PVTv2-B2 as a feature extraction network, input the environment image I to be detected into the feature extraction network to obtain multi-scale features Where i represents the number of layers in the feature extraction network, H and W are the length and width of the feature, respectively, and Ci represents the channel dimension of the feature in the i-th layer of the feature extraction network, The dimension of the vector representing the feature is H / 2i×W / 2i×Ci;

[0084] The second extraction unit will Input MRFB module and get

[0085] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A camouflaged target detection method based on subject and boundary perception, characterized in that: include: Step S1, obtaining an environment image to be detected; Step S2, preprocessing the environment image to be detected; Step S3, inputting the preprocessed environment image to be detected into the camouflage detection model to perform camouflage target detection; wherein the camouflage detection model includes: a body-boundary exploration module RBE, a body enhancement module REM and a boundary optimization module BRD; Step S2 includes: Step S21: select PVTv2-B2 as the feature extraction network, input the environment image I to be detected into the feature extraction network to obtain multi-scale features Where i represents the number of layers in the feature extraction network, H and W are the length and width of the feature, respectively, and C i represents the channel dimension of the features in the i-th layer of the feature extraction network, The dimension of the vector representing the feature is H / 2 i ×W / 2 i ×C i ; Step S22: Input MRFB module and get In step S3, I i 'Enter RBE to explore subject information and boundary information Will I i 'Aggregate with the output of the previous REM layer and and the output of the previous BRD are input into REM, wherein the fourth-layer REM has no previous BRD and previous REM, and only I'4 and As input; after each layer of REM, a BRD is set to enhance the boundary information again, and the output of each BRD passes through a segmentation head to obtain the prediction map and Supervised by region binary truth map, It is supervised by the boundary binary truth map; Total loss function of the camouflage detection model for: in, is the binary truth map of boundary information, is the Dice loss, is the mixed loss function, It consists of weighted BCE loss and weighted IoU loss. Among them, P represents the prediction graph, Represents the region binarization truth map.

2. A disguised target detection device based on subject and boundary perception for implementing the disguised target detection method based on subject and boundary perception as claimed in claim 1, characterized in that: include: An acquisition module, used for acquiring an environment image to be detected; A preprocessing module, used for preprocessing the environment image to be detected; The detection module is used to input the preprocessed environment image to be detected into the camouflage detection model for camouflage target detection; wherein the camouflage detection model includes: a body-boundary exploration module RBE, a body enhancement module REM and a boundary optimization module BRD.

Images