Detection method of camouflage object

Abstract

The invention discloses a detection method of a camouflage object. The detection method comprises the following steps: acquiring picture information of the camouflage object output by a preset model; acquiring a preset loss function set; preprocessing the preset loss function set to obtain a target loss function; and detecting the camouflage object based on the target loss function and the picture information of the camouflage object output by the preset model.

Description

Technical Field

[0001] The invention belongs to the technical field of image processing, and in particular relates to a method for detecting camouflaged objects. Background Art

[0002] Camouflage poses a great challenge to camouflaged object detection (COD). The key to camouflaged object detection is to accurately identify and segment camouflaged objects that are similar to the background.

[0003] The detection of disguised objects consists of a deep learning architecture and a loss function. In the prior art, although significant progress has been made in the deep learning architecture of disguised object detection, less attention has been paid to optimizing the loss function, which leads to class imbalance when detecting disguised objects, and thus results in less than ideal detection performance. Summary of the invention

[0004] The purpose of the embodiments of the present invention is to provide a method for detecting a disguised object, so as to solve the technical problem in the prior art that when detecting the performance of the disguised object, there is a category imbalance, which leads to unsatisfactory detection performance.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is to provide a method for detecting a camouflaged object, which is characterized by comprising:

[0006] Obtaining image information of the camouflaged object output by a preset model;

[0007] Get the preset loss function set;

[0008] Preprocessing the preset loss function set to obtain a target loss function;

[0009] The disguised object is detected based on the target loss function and the image information of the disguised object output by the preset model.

[0010] Preferably, the steps before obtaining the preset loss function set are:

[0011] The loss function of the COD (coefficient of determination) model is decoupled to obtain the preset loss function set.

[0012] Preferably, the step of obtaining the image information of the camouflaged object output by the preset model includes:

[0013] The image information of the camouflaged object is input into the preset model and then processed to obtain the image information of the camouflaged object output by the preset model.

[0014] Preferably, the specific steps of processing after inputting the preset model are:

[0015] The image information of the camouflaged object is input into a search module for search processing, and the first image information is output; the search module uses a pyramid visual transformer (PVT) as a backbone network;

[0016] Input the first picture information into an enlargement module for enlargement processing, and output to obtain second picture information;

[0017] The second image information is input into a recognition module for recognition processing, and the third image information is outputted. The third image information is used as the image information of the camouflaged object output by the preset model.

[0018] Preferably, the preset loss function set includes a binary cross entropy loss function, an intersection-over-union loss function and a Dice similarity coefficient loss function.

[0019] Preferably, the preset model is a SARNet model.

[0020] Preferably, the preset loss function set is preprocessed to obtain the objective function, and the specific steps are:

[0021] The preset loss functions included in the preset loss function set are weighted respectively, and then linearly combined to obtain the objective function.

[0022] Preferably, the detecting of the disguised object based on the target loss function and the image information of the disguised object output by the preset model is specifically as follows:

[0023] The image information of the camouflaged object output by the preset model is input into the target loss function to detect the camouflaged object.

[0024] The present invention has the following beneficial effects:

[0025] The present invention provides a method for detecting a disguised object. After preprocessing a preset function set, a target loss function is obtained to solve the problem of class imbalance and improve detection performance. The target loss function is combined with the image information of the disguised object obtained by obtaining the output of the preset model to detect the disguised object accordingly, thereby significantly improving the detection performance.

[0026] In addition to the above-described purposes, features and advantages, the present invention has other purposes, features and advantages. The present invention will be further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The drawings constituting a part of this application are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0028] Figure 1 It is a specific structural schematic diagram of the detection method of the present application;

[0029] Figure 2 It is a flowchart of this application.

[0030] Figure 3 It is a schematic diagram of a process of obtaining the image information of the camouflaged object output by the preset model provided in a preferred embodiment of the present application.

[0031] Figure 4 is a schematic diagram showing a visual comparison of the results of the present application with those of the advanced methods. DETAILED DESCRIPTION

[0032] The embodiments of the present invention are described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.

[0033] The detection of camouflaged objects is crucial. In the military, it can help discover hidden military targets and enhance national defense security. In the field of security, it can protect important facilities and public safety. In ecological protection, it helps to study rare species and monitor ecology. Its wide application demonstrates its key value in different fields.

[0034] The key point of detecting camouflaged objects lies in the post-processing of images, which is mainly processed by deep learning architecture and loss function. The loss function plays a key role in image processing. It measures the difference between the model prediction and the actual result, and guides the model to optimize parameters to improve the accuracy and quality of image processing. In the existing technology, people's research direction is mainly to improve the performance of deep learning architecture, ignoring the optimization of loss function, which leads to certain errors in image processing.

[0035] In order to solve the above problems, the present application proposes Figure 1 The figure shows an overall process diagram for optimizing the loss function, thereby reducing the error in image processing and improving the accuracy of image detection.

[0036] First, obtain a picture of the camouflaged object to be detected and input the picture into the SARNet model. The specific processing process is as follows: first input it into the search module in the model for search processing, then input the processed content into the magnification module for magnification processing, then input the amplified content into the recognition module for corresponding recognition, and then input the recognized content into the optimized wBID loss function for optimization training, thereby improving the accuracy of detection.

[0037] In order to make it easier to understand the content of this application, according to Figure 2 The specific contents of this application are described in detail.

[0038] Step S101: obtaining the image information of the camouflaged object output by a preset model;

[0039] The preset model described in this application is the SARNet model. In image processing, SARNet can extract multi-level features, use spatial attention to enhance features, and assist in image segmentation, such as dividing object boundaries, etc., which has a significant effect. In the field of images, a camouflaged object refers to an object that hides itself by being similar to the background through color, texture, etc. Therefore, the image information of the camouflaged object mentioned here in this application can be simply understood as a picture taken of the camouflaged object. Here, the way to shoot the camouflaged object can be to use a mobile phone, a camera, or a special image detection device to shoot. There is no specific restriction here. According to the specific usage scenario, the technician can set it by himself.

[0040] Combination Figure 1 , the above process is explained in detail. The SARNet model in this application mainly includes a search module, an amplification module, and a recognition module.

[0041] Here is a brief description of the role of the above modules in image processing:

[0042] Search module. In this model, the search part uses the Pyramid Vision Transformer (PVT) as the backbone network to extract multi-level features. It generates high-resolution feature maps through attention-based feature search while reducing memory overhead. These feature maps are processed by the Object Area Amplification (OAA) and Image Foreground-Background Conversion (FGC) modules and then passed to the amplification and recognition parts to ensure efficient connectivity.

[0043] Amplification module. The amplification part consists of four OAA (Occlusion-Aware Attention) modules and one CBR (Convolution-Batch Normalization-ReLU) module, which promotes the fusion of features across layers and adjacent layers, and solves the problem that the search stage may affect recognition due to insufficient features. This part assists in the fine segmentation of the recognition part by scaling the feature map, while ensuring strong connectivity between the search and recognition components.

[0044] Recognition module. The recognition part contains four image foreground-background conversion modules (FFGC (Fully-Featured Graph Convolution) and EFGC (Enhanced Foreground-Background Classification), which are variants of the figure-ground conversion (FGC) module. FFGC emphasizes foreground information, while EFGC focuses on background edge details. The FGC module dynamically selects the focus of foreground or background, and processes the current level features, deep features, and lateral predictions. This continuous optimization process improves segmentation accuracy and reduces errors in foreground and background division.

[0045] In this application, before obtaining the image information of the camouflaged object output by the preset model, the image information of the camouflaged object needs to be processed first, and the specific process of the processing is as follows;

[0046] The image information of the camouflaged object is input into the preset model and then processed to obtain the image information of the camouflaged object output by the preset model.

[0047] Combination Figure 1 The specific process of the above processing is described as follows:

[0048] Step S301: input the image information of the camouflaged object into a search module for search processing, and output first image information; the search module uses a pyramid visual Transformer (PVT) as a backbone network;

[0049] This step mainly uses the picture information of the camouflaged object to generate a high-resolution feature map based on the attention-based feature search in the search part. In order to achieve this effect, the search part of this application uses the Pyramid Vision Transformer (PVT) as the backbone network to extract multi-level features. The Pyramid Vision Transformer (PVT) has the following advantages as a backbone network for feature extraction: it can extract features in layers, adapt to targets of different scales, and has a global receptive field, captures long-distance dependencies, and assists image processing tasks. The first picture information mentioned in this application can be understood as: the picture information of the camouflaged object, which is the picture information obtained after the feature extraction process in the search module.

[0050] Step S302: inputting the first picture information into an amplification module for amplification processing, and outputting second picture information;

[0051] The magnification module in this application is mainly composed of four OAA (Occlusion-Aware Attention) modules and one CBR (Convolution-Batch Normalization-ReLU) module. In the image field, the OAA module is a component for optimizing feature extraction or enhancement.

[0052] In target detection or multi-target tracking tasks, the output of the OAA module will be used as the input feature for subsequent classification, detection or Re-ID (Re-Identification) tasks, helping the model to more accurately detect potentially occluded targets and improve the model's ability to recognize and track targets.

[0053] In the image field, the CBR module refers to case-based reasoning, which processes current image tasks by referring to past similar image cases, assisting in image analysis and processing, etc.

[0054] Therefore, the second image information in the present application can be understood as the image information obtained after the first image information is input into the enlargement module and processed by the OAA module or the CBR module in the enlargement module.

[0055] Step S303: input the second picture information into a recognition module for recognition processing, output to obtain third picture information, and use the third picture information as the picture information of the camouflaged object output by the preset model.

[0056] The recognition module in this application mainly includes four image foreground-background conversion modules (FFGC and EFGC), which are variants of the graph-ground conversion (FGC) module. FFGC (Fully-Featured Graph Convolution), this algorithm may use graph convolution to process pixel relationships in the image, and fully explore the structural information of the image for tasks such as feature extraction or segmentation. EFGC (Foreground-Background Classification) is a basic and important task in image processing. Its purpose is to accurately separate the foreground objects in the image (such as people, vehicles, animals and other objects of main concern) from the background (such as the sky, ground, buildings and other relatively unimportant parts). For example, in video surveillance, the FGC technology can quickly identify moving objects (foreground) so that they can be tracked and analyzed; in image editing, after separating the foreground and background, the foreground can be easily cut out, the background can be replaced, and other operations can be performed. There are many ways to implement FGC, and the most common one is to distinguish based on features such as color, texture, and edges. For example, the color of the foreground object is usually quite different from the background, or the edge of the foreground object is obvious. These features can be used to design algorithms to segment the image.

[0057] Therefore, the third picture information in the present application can be understood as the picture information obtained after the second picture information is input into the recognition module and processed by the FFGC or EFGC module in the recognition module.

[0058] Step S102: Obtain a preset loss function set;

[0059] In image processing, the loss function measures the difference between the predicted image and the real image, and is used to guide model training so that the model outputs more accurate images. This application focuses on camouflaged object detection (COD). In the field of images, COD is a specific model (assuming it is a Center-Oriented Detection model), and its loss function is critical. It is used to measure the degree of difference between information such as the target center position predicted by the model and the real annotation. By minimizing this loss, the model is prompted to locate the target center more accurately and improve detection accuracy.

[0060] Since the loss function used in this application is not a commonly used function. Therefore, before obtaining the preset loss function set, this application needs to decouple the loss function of the COD model to obtain the preset loss function set. The preset loss function set generally includes a binary cross entropy loss function, an intersection over union (IoU) loss function, and a Dice (Dice coefficient) loss function.

[0061] Here, the above three functions are briefly introduced:

[0062] Binary cross entropy loss function: It is widely used in binary classification problems in machine learning. It calculates the loss value by comparing the predicted probability with the true label. When the predicted probability is close to the true label (0 or 1), the loss decreases. It is often used to distinguish between two types of targets in image classification, helping the model optimize parameters to improve classification accuracy.

[0063] The intersection-over-union (IoU) loss function is commonly used in object detection tasks to measure the degree of overlap between the predicted box and the true box. The loss is defined by calculating the difference between the intersection and union ratios, prompting the model to optimize the predicted box position and size to improve the accuracy and precision of object detection.

[0064] The DICE loss function is mainly used in the field of image segmentation. It is based on the similarity measurement of two sets and measures the overlap between the segmentation result and the true annotation by calculating the ratio of the intersection and the union. It can effectively optimize the image segmentation model and improve the accuracy and quality of segmentation.

[0065] Step S103: preprocessing the preset loss function set to obtain a target loss function;

[0066] In this application, after obtaining the above-mentioned preset loss function set, it is necessary to preprocess it. The preprocessing here mainly refers to the preset loss functions of the blocks in the preset loss function set, which are weighted respectively, and then linearly combined to obtain the objective function. The objective function here is the wBID loss function. In other words, in this application, the binary cross entropy loss function, the intersection over union (IoU) loss function and the Dice (Dice coefficient) loss function are weighted respectively, and then linearly combined to obtain the wBID loss function.

[0067] In this application, the specific process of preprocessing the loss function in the loss function set is as follows:

[0068] Weighted Binary Cross Entropy Loss. Weighted Binary Cross Entropy Loss (wBCE Loss) assigns different weights to disguised targets (positive samples) and backgrounds (negative samples) to prioritize identifying disguised targets. This enhances the accuracy of pixel-level classification, which is especially important when dealing with class imbalance problems in disguised target detection. The formula for wBCE Loss is:

[0069]

[0070] In the formula, N represents the total number of samples in the training dataset and is used for loss normalization. i represents the true label of the i-th sample, where represents the positive sample (disguised target), represents negative samples (background), and and Represent the model’s predicted probabilities for the positive and negative classes, respectively. w pos and w neg are the weights of positive samples and negative samples respectively. In the camouflaged target detection task, w pos Usually set high to emphasize learning of camouflaged targets, while w neg Set it low to reduce the impact of background on model training.

[0071] Weighted IoU loss function. Weighted intersection over union loss (wIOU Loss) is used to evaluate the overlap between the predicted disguised target and the real disguised target, similar to wBCE loss. By adjusting the weights to cope with class imbalance, wIOU improves the model's ability to identify disguised targets, thereby improving the performance of disguised target detection. The formula for wIOU Loss is:

[0072]

[0073] It should be noted that the meaning of each term in this formula is the same as that of each term in formula (1), so the meaning of each term in the formula will not be described in detail here. The same applies to formula (3).

[0074] Weighted Dice loss. Weighted Dice loss is a loss function widely used in image segmentation that evaluates the similarity between the predicted region and the true region. Similar to weighted IoU loss, it evaluates segmentation accuracy by comparing overlapping regions. By weighting positive and negative samples, it solves the sample imbalance problem in camouflage detection and prioritizes accurate segmentation of camouflaged targets. This improves its robustness under various camouflage types and complex backgrounds. The formula for weighted Dice loss is:

[0075]

[0076] In summary, wBCE loss improves pixel classification accuracy and helps identify camouflaged targets; wIoU loss enhances positioning accuracy and makes predictions more consistent with true labels; wDice loss focuses on shape matching, improving segmentation accuracy and the robustness of camouflaged target detection. The linear combination of these three loss functions can fully guide model training, thereby improving the performance of camouflaged target detection tasks. The formula is:

[0077]

[0078] Step S104: Detecting the disguised object based on the target loss function and the image information of the disguised object output by the preset model.

[0079] In the present application, the main implementation process of this step is to input the image information of the camouflaged object output by the preset model into the target loss function, so as to detect the camouflaged object and identify the camouflaged object in the image information. Figure 1 The specific process of implementing the above steps is described as follows:

[0080] SARNet has five output heads, which come from the search part and the recognition part. The output feature map of the search part is processed by the target area expansion module and the image foreground-background conversion module, and then passed to the expansion part and the recognition part to establish an effective connection. The five output predictions are p = {p i |i=1,2,3,4,5} and resized to the same resolution as the input to obtain the final model output {P i |i=1,2,3,4,5}, where P1 is the final prediction result. In this application, the wBID loss function is applied to each output head, and l is defined according to the model structure and the definition formula of each basic loss function. bce , l iou and l dice as follows:

[0081]

[0082]

[0083]

[0084] Where L represents the total number of pixels, G represents the true label of the camouflaged target, and δ is the indicator function. The loss function used is as follows:

[0085] l fm =αl bce +βl iou +γl dice (8)

[0086]

[0087] In the loss function, l fm L is the foreground matching loss, which is an important indicator for measuring the prediction accuracy of the model for the camouflaged target area. It consists of binary cross entropy loss, intersection-over-union loss and Dice loss. Each part is weighted by weight coefficients α, β, and γ, where α = 2, β = 1, and γ = 1; L overall is the total loss, which is based on l fmA high-order polynomial combination, mainly used to optimize the overall performance of the model, L overall Yes fm The weighted sum of multiple powers of expresses the comprehensive optimization of the model for the foreground and background areas.

[0088] In order to verify the implementation effect of this application, the following experimental data is provided here.

[0089] Evaluation Metrics:

[0090] For the camouflaged target detection task in this application, six commonly used evaluation indicators are used, including S-measure (S m ), Mean F-measure Weighted F-measure Mean E-measure Max E-measure and Mean Absolute Error Among them, S-measure aims to evaluate the structural information of the prediction results; E-measure focuses more on pixel-level matching and image-level statistics; and F-measure emphasizes the overall measurement of the precision and recall of the predicted image. Mean Absolute Error calculates the element-by-element difference.

[0091] Experimental details: This application uses PyTorch to implement the model and conducts six experiments, including SINet[1], PFNet

[13] , SINet-V2[4], ZoomNet[3], FSPNet[2], and SARNet[6]. Among them, PFNet, ZoomNet, and FSPNet use official test results, while SINet, SINet-V2, and SARNet use official source code. Training and testing are performed on a machine equipped with an Intel(R) Xeon(R) Platinum 8352V CPU@

[0092] The test was performed on a 36-core PC running at 2.10GHz (64GB of memory) and four NVIDIA GeForce RTX 4090 GPUs (24GB each).

[0093] During training, this application used 4040 camouflaged target images from COD10K. The input images were resized to 384×384 and data augmentation was performed, including random horizontal flipping and color jittering. The optimizer used stochastic gradient descent (SGD) with momentum 0.9 and weight decay 5×10-4. The batch size for training was set to 16, and training was performed for 100 epochs with a base learning rate of 0.001. During testing, the input images were resized to 384×384 for inference and then resized back to the original size.

[0094] No post-processing was performed on the final output. Performance evaluation was conducted on a device equipped with an Intel(R) Core(TM) i9-13900HX CPU and an NVIDIA GeForce RTX 4050 laptop GPU.

[0095] Experimental results:

[0096] In order to verify the effectiveness of the network proposed in this application, this application compares it with some existing advanced methods using three benchmark datasets and six evaluation indicators. Existing advanced methods include SINet[1], PFNet

[13] , SINet-V2[4], ZoomNet[3], FSPNet[2], and SARNet[6]. To ensure a fair comparison, this application downloaded the results of these state-of-the-art models and evaluated them using the same evaluation tool. As shown in Table 1,

[0097]

[0098] Table 1 shows the best, second best and third best performance methods in bold black font.

[0099] As shown in Table 1, in Table 1, the model of this application outperforms other methods on all three datasets, showing excellent performance. FSPNet usually ranks second in various evaluation indicators, followed by ZoomNet on CAMO and COD10K datasets, and SARNet on NC4K dataset. Compared with FSPNet, wBID-SNet improves S-measure by 0.579%, F-measure by 2.512%, and E-measure by 1.473% on average in all datasets. Similarly, compared with ZoomNet, wBID-SNet has an average improvement of 3.600%, 5.577%, and 3.909% on S-measure, F-measure, and E-measure on CAMO and COD10K datasets, respectively. In the NC4K dataset, SARNet has an average improvement of 1.362%, 3.255%, and 1.656% on S-measure, F-measure, and E-measure, respectively, compared with wBID-SNet of this application. These results highlight the superior performance and versatility of wBID-SNet in this application for camouflaged object detection.

[0100] like Figure 4 As shown, the visual results of the detection method provided by the present application (the third row) are compared with other advanced methods. It should be noted that Ours in the third row of the figure represents the detection method of the present application. It can be directly observed from the displayed results that the method proposed in the present application is intuitively compared with other representative competitive methods in various typical scenes, including small objects (first and second columns), large objects (third and fourth columns), multiple objects (fifth and sixth columns), occluded objects (seventh and eighth columns) and targets with uncertain boundaries (ninth and tenth columns). For small objects, the model of the present application can locate them more accurately; for large objects, occluded objects and camouflaged targets with uncertain boundaries, the model of the present application can still segment their edges more carefully. In addition, it can be seen from the figure that in the detection of multiple camouflaged targets, the visual effects of all the most advanced methods are poor, while the method of the present application performs well in the right image (sixth column) of multiple camouflaged targets, close to the true annotation, but performs poorly in the left image (fifth column) compared with other most advanced methods.

[0101] In this application, the impact of loss functions on camouflaged target detection tasks is discussed. This application finds that with the development of camouflaged target detection networks based on deep learning, the loss function has gradually evolved from the initial use of simple cross-entropy loss to the introduction of IoU loss and Dice loss according to task requirements, and the introduction of weights and specific loss functions to assist enhancement modules according to different network characteristics, forming a complex loss function. This application decouples these complex loss functions into three basic loss functions: Binary Cross-Entropy Loss, IoU Loss, and Dice Loss, and linearly combines them after weighting to fully exert their advantages in classification accuracy, positioning accuracy, and shape matching. This application applies these loss functions to the SARNet architecture to obtain wBID-SNet, and compares it with other high-performance camouflaged target detection networks on CAMO, COD10K, and NC4K datasets. The results show that the improved camouflaged object detection method of this application outperforms other advanced methods on all three datasets, which not only helps the subsequent research on camouflaged target detection tasks, but also improves its detection performance.

[0102] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for detecting a camouflaged object, characterized in that: include: Obtaining image information of the camouflaged object output by a preset model; Get the preset loss function set; Preprocessing the preset loss function set to obtain a target loss function; The disguised object is detected based on the target loss function and the image information of the disguised object output by the preset model.

2. The detection method according to claim 1, characterized in that: The steps before obtaining the preset loss function set are: The loss function of the COD (coefficient of determination) model is decoupled to obtain the preset loss function set.

3. The detection method according to claim 1, characterized in that: The step of obtaining the image information of the camouflaged object output by the preset model includes: The image information of the camouflaged object is input into the preset model and then processed to obtain the image information of the camouflaged object output by the preset model.

4. The detection method according to claim 3, characterized in that: The specific steps of processing after inputting the preset model are: The image information of the camouflaged object is input into a search module for search processing, and the first image information is output; the search module uses a pyramid visual transformer (PVT) as a backbone network; Input the first picture information into an enlargement module for enlargement processing, and output to obtain second picture information; The second image information is input into a recognition module for recognition processing, and the third image information is outputted. The third image information is used as the image information of the camouflaged object output by the preset model.

5. The detection method according to claim 1, characterized in that: The preset loss function set includes a binary cross entropy loss function, an intersection-over-union loss function, and a Dice similarity coefficient loss function.

6. The detection method according to claim 1, characterized in that: The preset model is the SARNet model.

7. The detection method according to claim 1, characterized in that: The preset loss function set is preprocessed to obtain the objective function, and the specific steps are: The preset loss functions included in the preset loss function set are weighted respectively, and then linearly combined to obtain the objective function.

8. The detection method according to claim 4, characterized in that: The detecting of the camouflaged object based on the target loss function and the image information of the camouflaged object output by the preset model is specifically as follows: The image information of the camouflaged object output by the preset model is input into the target loss function to detect the camouflaged object.

Images