A method, system, device and medium for identifying adaptive camouflaged targets in complex scenes

Abstract

The present invention discloses a method, system, device and medium for identifying adaptive camouflaged targets in complex scenes, and relates to the field of computer vision technology of artificial intelligence. The method comprises: inputting the acquired image to be tested into a target recognition model for processing to obtain a target recognition result; the target recognition model is constructed based on an interactive guidance network and a saliency perception collaborative optimization loss function; the interactive guidance network includes an adaptive multi-scale feature enhancer, a probabilistic context aggregation module and a cross-scale attention fusion module connected in sequence; the saliency perception collaborative optimization loss function is composed of five weighted parts, each of which is a binary cross entropy loss, a joint intersection loss, a mean square error loss, a structural similarity index metric and a mean absolute error loss. The present invention can improve the intelligence level of public security monitoring and enhance the early warning capability of potential threat targets.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence computer vision technology, and in particular to a method, system, device and medium for identifying adaptive camouflaged targets in complex scenes. Background Art

[0002] With the rapid development of smart cities and public security, smart policing technology has become a core pillar of modern security systems. However, in complex scenarios (such as densely populated areas, nighttime environments, or when camouflaged targets interfere), traditional target detection and recognition methods often struggle to detect camouflaged targets (such as camouflaged vehicles, concealed weapons, and disguised objects) in high-resolution images. These targets often blend seamlessly with the background through color, texture, or behavioral patterns, making conventional detection algorithms susceptible to interference, resulting in reduced recognition accuracy and impacting the real-time and reliability of policing decisions.

[0003] Therefore, the main issues currently include the following:

[0004] 1. Insufficient recognition robustness in complex scenarios: When the target and background have highly similar visual features, traditional detection methods struggle to effectively distinguish between the target and background interference. While deep learning-based detection frameworks have improved feature extraction capabilities, they still suffer from significant false positives and missed detections when handling complex situations such as multi-scale targets, dynamic lighting changes, and partial occlusion. In particular, the recognition accuracy of existing methods drops significantly when the target employs active camouflage strategies.

[0005] 2. Inadequate recognition of small targets and fine features: Traditional methods are ineffective at detecting camouflaged targets occupying smaller areas within an image. Due to the limited pixel information available, key features of these targets are easily overlooked or weakened in conventional processing. Furthermore, traditional methods are incapable of capturing edge details, making it difficult to accurately identify boundary features of even finely camouflaged targets. This results in inaccurate target outline extraction, directly impacting subsequent analysis and prediction.

[0006] 3. Computational efficiency and real-time performance are difficult to balance: High-resolution image processing places extremely high demands on computing resources. Existing detection algorithms generally suffer from low computational efficiency when processing ultra-high-resolution inputs. Especially when deployed on edge computing devices, the real-time performance of these methods often fails to meet practical application requirements. As a result, the balance between model lightweighting and detection accuracy has yet to be effectively resolved.

[0007] 4. Limitations in scenario adaptability and generalization: Existing methods are highly dependent on the distribution of training data, resulting in poor performance in novel scenarios or with unseen camouflage types. Due to the diverse and dynamic nature of camouflage strategies in real-world scenarios, models trained on fixed datasets struggle to consistently maintain optimal recognition performance. Furthermore, the lack of an effective incremental learning mechanism makes it difficult for models to adapt to the continuous evolution of camouflage methods, leading to performance degradation after field deployment. Summary of the Invention

[0008] The purpose of the present invention is to provide a method, system, device and medium for identifying adaptive camouflaged targets in complex scenarios, aiming to solve or improve at least one of the above-mentioned technical problems.

[0009] To achieve the above object, the present invention provides the following solutions:

[0010] A method for identifying adaptively camouflaged targets in complex scenarios, comprising:

[0011] Acquire an image to be tested; the image to be tested is a captured image with a complex scene;

[0012] The image to be tested is input into a target recognition model for processing to obtain a target recognition result; the target recognition model is constructed based on an interactive guidance network and a saliency-aware collaborative optimization loss function; the interactive guidance network includes an adaptive multi-scale feature enhancer, a probabilistic context aggregation module, and a cross-scale attention fusion module connected in sequence; the saliency-aware collaborative optimization loss function consists of five weighted parts, namely binary cross entropy loss, joint intersection loss, mean square error loss, structural similarity index measurement, and mean absolute error loss.

[0013] Optionally, the image to be tested is input into a target recognition model for processing to obtain a target recognition result. The specific process includes:

[0014] The ResNet-50 feature encoder is used to extract multi-scale spatial features of the image to be tested, and multi-scale features from low-level features to high-level features are obtained. 、 、 、 and Among them, the characteristics and For low-level features, features 、 and For advanced features;

[0015] The multi-scale features 、 、 、 and Input the adaptive multi-scale feature enhancer to enhance visual information and generate visual enhancement features 、 、 、 and ;

[0016] For the multi-scale features 、 、 、 and And the visual enhancement features 、 、 、 and Perform feature concatenation and channel compression transformation to generate probability map features , local graph features and aggregate features , and input the generated features into the probability context aggregation module to calculate the probability context aggregation features ;

[0017] Aggregate the probabilistic context features The cross-scale attention fusion module is input for prediction to obtain the final target recognition result.

[0018] Optionally, the operation process of the adaptive multi-scale feature enhancer specifically includes:

[0019] Step 1: Input features Perform convolution calculations and cross-kernel operations separately to obtain multi-branch output features with different receptive fields 、 、 and :

[0020] ,

[0021] in, Represents the multi-scale spatial features extracted by ResNet-50, Indicates that the convolution kernel size is The convolution operation, stands for cross-kernel operation;

[0022] Step 2: Feature concatenation and nonlinear correction units along the channel dimension are applied to the multi-branch output features 、 、 and , thereby generating channel weighted features :

[0023] ,

[0024] in, represents the feature concatenation operation in the channel dimension, Represents the nonlinear correction unit operation composed of Conv, BN and ReLU;

[0025] Step 3: Combine the channel squeeze and excitation module with the channel multiplication operation and apply it to the channel weighted features , thereby obtaining channel fusion features :

[0026] ,

[0027] in, represents the channel squeezing and excitation module, Represents channel multiplication operation;

[0028] Step 4: Residual connection operation, element-wise addition operation and nonlinear correction unit are applied to the input features and channel fusion features , to generate visual enhancement features :

[0029] ,

[0030] in, Represents residual learning, + represents element-by-element addition operation, Represents the nonlinear correction unit operation.

[0031] Optionally, the cross-core operation process specifically includes:

[0032] Step 1: Using the cross operation involves the input features Horizontal and vertical convolution operations are applied sequentially to generate features :

[0033] ,

[0034] in, and Represent the horizontal and vertical convolution operations respectively, and Represents the convolution kernel matrix;

[0035] Step 2: Generate features Perform matrix multiplication and softmax function normalization to obtain the generated features Horizontal attention score :

[0036] ,

[0037] in, represents matrix multiplication, represents the transpose transformation, represents the feature dimension, 、 and represents the weights used to generate query, keyword, and value vectors in the horizontal attention branch, represents the softmax normalization function;

[0038] Step 3: Generate features Perform matrix multiplication and softmax function normalization to obtain the generated features Vertical attention score :

[0039] ,

[0040] in, represents matrix multiplication, represents the transpose transformation, represents the feature dimension, , and represents the weights used to generate query, keyword, and value vectors in the vertical attention branch, represents the softmax normalization function;

[0041] Step 4: Score the horizontal attention and vertical attention score Use element-wise addition, Hadamard product, and nonlinear correction unit operations to generate branch output features , which is expressed as follows:

[0042] ,

[0043] Among them, + represents element-by-element addition operation, represents the Hadamard product operation, represents the softmax normalization function, Indicates use Sub-nonlinear correction unit operation.

[0044] Optionally, the computation process of the probabilistic context aggregation module specifically includes:

[0045] Step 1: In the probability map feature Calculate the probability offset and generate probability predictions for foreground, background and confusion regions respectively 、 and :

[0046] ,

[0047] in, Represents the global threshold, the initial value is set to 0.5, represents metric normalization, represents the bilinear interpolation operation, represents the Sigmoid function, Represents convolution operation and channel compression transformation;

[0048] Step 2: In the local graph features Calculate the local probability offset and generate local foreground probability predictions respectively and local background probability prediction , respectively expressed as:

[0049] ,

[0050] in, Represents the local threshold, the initial value is set to 0.5, represents metric normalization, represents the bilinear interpolation operation, represents the Sigmoid function, Represents convolution operation and channel compression transformation;

[0051] Step 3: Global Probability Prediction 、 and and local probability predictions and Channel concatenation and nonlinear correction are performed to generate fused probability features , expressed as:

[0052] ,

[0053] in, represents the feature concatenation operation in the channel dimension, Represents the nonlinear correction unit operation;

[0054] Step 4: Aggregate features and fusion probability features Perform scale reconciliation, feature reshaping, and probability prediction, and use Hadamard product to calculate context features :

[0055] ,

[0056] in, represents the bilinear interpolation operation, represents feature reshaping, Represents tensor rearrangement by dimension transformation, Represents convolution operation and channel compression transformation;

[0057] Step 5: Refine the salient objects by using similarity matrix calculation and value calculation to obtain attention scores and refinement features :

[0058] ,

[0059] in, represents matrix multiplication, represents the transpose transformation, represents the feature dimension, represents the softmax normalization function, , and Both represent convolution kernel matrices;

[0060] Step 6: Use spatial attention mechanism and nonlinear transformation to capture feature dependencies in spatial dimensions to obtain attention features as follows:

[0061] ,

[0062] in, represents the Hadamard product operation, represents the ReLU function, represents the Sigmoid function, and Represents the convolution kernel matrix;

[0063] Step 7: Aggregate features of input and attention features The fused and refined features are normalized by the nonlinear correction unit and the final output is The mathematical expression is:

[0064] ,

[0065] in, represents the feature cascade operation on the channel dimension, Indicates the use of cubic nonlinear correction unit operation.

[0066] Optionally, the operation process of the cross-scale attention fusion module specifically includes:

[0067] Step 1: Input features Use transposed convolution and pixel shuffling for upsampling, followed by multi-scale convolution operations to generate local features 、 and :

[0068] ,

[0069] in represents an upsampling operation using transposed convolution and pixel shuffling, Indicates size Multi-scale convolution kernel ;

[0070] Step 2: Local features 、 and Perform channel dimension feature cascade and refine the cascade features through the nonlinear correction unit to obtain refined features :

[0071] ,

[0072] Step 3: Refine the features The horizontal attention score is generated by matrix multiplication and normalization function calculation and horizontally weighted features :

[0073] ,

[0074] in, , and represents the weights used to generate query, keyword, and value vectors in the horizontal attention branch, represents matrix multiplication, represents the transpose transformation, represents the feature dimension, represents the softmax normalization function;

[0075] Step 4: Refine the features The vertical attention score is generated by matrix multiplication and normalization function calculation and vertical weighted features :

[0076] ,

[0077] in , and represents the weights used to generate query, keyword, and value vectors in the vertical attention branch, represents matrix multiplication, represents the transpose transformation, represents the feature dimension, represents the softmax normalization function;

[0078] Step 5: Horizontally weight features and vertical weighted features are fused, rectified, and channel-compressed to obtain the final output features as follows:

[0079] ,

[0080] Among them, + represents element-by-element addition operation, Indicates the use of a quadratic nonlinear correction unit, Represents the channel compression transform.

[0081] Optionally, the formula of the saliency-aware collaborative optimization loss function is expressed as:

[0082] ,

[0083] in, Co-optimize loss function for saliency perception; 、 and is the weight coefficient, set to , and ; For the Binary cross entropy loss at the stage Combined loss sum; For the The mean squared error loss of the stage Structural similarity index sum; For the The mean absolute error loss of the stage .

[0084] The present invention also provides a recognition system for adaptively camouflaged targets in complex scenarios, comprising:

[0085] An image acquisition unit, configured to acquire an image to be tested; the image to be tested is a captured image with a complex scene;

[0086] The target recognition unit is used to input the image to be tested into the target recognition model for processing to obtain a target recognition result; the target recognition model is constructed based on an interactive guidance network and a saliency-aware collaborative optimization loss function; the interactive guidance network includes an adaptive multi-scale feature enhancer, a probabilistic context aggregation module, and a cross-scale attention fusion module connected in sequence; the saliency-aware collaborative optimization loss function consists of five weighted parts, namely binary cross entropy loss, joint intersection loss, mean square error loss, structural similarity index measurement, and mean absolute error loss.

[0087] The present invention also provides an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute the above-mentioned method for identifying adaptive camouflaged targets in complex scenarios.

[0088] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-mentioned method for identifying adaptive camouflaged targets in complex scenarios.

[0089] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0090] The present invention discloses a method, system, device, and medium for identifying adaptively camouflaged targets in complex scenarios. The method comprises inputting an acquired image to be tested into a target recognition model for processing to obtain a target recognition result. The target recognition model is constructed based on an interactive guidance network and a saliency-aware collaborative optimization loss function. The interactive guidance network includes an adaptive multi-scale feature enhancer, a probabilistic context aggregation module, and a cross-scale attention fusion module connected in sequence. The saliency-aware collaborative optimization loss function consists of five weighted components, namely, a binary cross-entropy loss, a joint intersection loss, a mean square error loss, a structural similarity index metric, and a mean absolute error loss. The present invention can improve the intelligence level of public security monitoring and enhance the early warning capability of potential threat targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0092] Figure 1 Flowchart of the method for identifying adaptive camouflaged targets in complex scenarios in this embodiment;

[0093] Figure 2 3 is a comparison chart of the significance prediction results using the method of the present invention and other methods in this embodiment. DETAILED DESCRIPTION

[0094] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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 making creative efforts are within the scope of protection of the present invention.

[0095] The purpose of the present invention is to provide a method, system, device and medium for identifying adaptive camouflaged targets in complex scenarios, aiming to solve or improve at least one of the above-mentioned technical problems.

[0096] 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.

[0097] This paper provides a method for adaptively recognizing camouflaged targets in complex scenarios. In terms of technical setup, considering the effectiveness of multi-feature interactive fusion for camouflaged target refinement and prediction, an interactive guidance network is first designed as a baseline for saliency inference and prediction. This network includes a ResNet-50 feature encoder to extract multi-scale spatial features and a layer-by-layer recursive fusion up-sampled feature decoder for detail enhancement and scale recovery. To generate coding features rich in visual enhancement information, an adaptive multi-scale feature enhancer is proposed and integrated into the interactive guidance network. The adaptive multi-scale feature enhancer dynamically captures long-range dependencies in the horizontal and vertical spaces through a cross-attention-guided multi-directional feature decomposition and coordinates feature responses across different receptive fields using a channel-space dual-path optimization mechanism. It employs a channel recalibration strategy driven by a squeezing and excitation module to address the imbalance in contributions between multi-scale features. Combined with residual learning to maintain gradient propagation efficiency, this enhances feature representation that balances both discriminative power and robustness in complex scenarios. The probabilistic context aggregation module explicitly models the semantic uncertainty of foreground, background, and confusion regions through probability-driven feature decoupling, and dynamically fuses multi-scale contextual information using local-global probability co-correction to resolve classification ambiguity caused by semantic fuzziness. It further optimizes feature discriminability through attention-guided cross-region correlation and captures long-range dependencies through spatial nonlinear transformation, achieving highly robust feature representation while ensuring computational efficiency. The cross-scale attention fusion module constructs a local detail-enhanced representation through upsampling-guided multi-scale feature reorganization and collaboratively models long-range dependencies in the horizontal and vertical spaces using a bidirectional attention mechanism. This module implements feature distillation through channel compression transformation, eliminating redundant responses while maintaining the integrity of multi-scale information, ultimately outputting fused features that are both spatially accurate and semantically consistent.

[0098] In terms of optimization settings, the present invention proposes a saliency-aware collaborative optimization loss function, which can refine the generated saliency prediction map while improving the accuracy of saliency prediction. The function consists of five weighted parts: binary cross entropy loss is combined with joint intersection loss to effectively improve the model's robustness to noise interference, complex background and category imbalance problems; mean square error loss is combined with structural similarity index measurement to simultaneously optimize pixel-level accuracy and structural similarity, thereby improving the accuracy and visual quality of the prediction results; mean absolute error loss can balance global and local information to generate more natural and accurate saliency maps. These five loss functions complement each other and can enhance the accuracy of the generated saliency prediction map while removing interference information and noise in the image. In terms of result comparison, the experimental results of the method proposed in this invention on nine public datasets show that the method of this invention is superior to other existing cutting-edge methods.

[0099] Therefore, if Figure 1 As shown in the figure, the method proposed in this invention uses ResNet-50 as a feature encoder to extract multi-scale spatial features, which are respectively denoted as 、 、 、 and ;in and It is a low-level feature that contains detailed target details and rich target location information, which is more conducive to the reasoning and prediction of salient target details and boundaries. 、 and As high-level features, they contain higher-level abstract semantic information, which can more effectively characterize the essential attributes and distinguishing features of the target category, thereby improving the ability to distinguish disguised targets. First, the present invention combines multi-scale features 、 、 、 and As the input of the adaptive multi-scale feature enhancer to generate a feature tensor containing rich visual enhancement information 、 、 、 and Second, multi-scale features 、 、 、 and and visual enhancement features 、 、 、 and Perform feature concatenation and channel compression transformation to generate probability map features , local graph features and aggregate features . Third, the generated probability map features , local graph features and aggregate features As the input of the probabilistic context aggregation module to calculate the fused probabilistic features , contextual features , refine features , attention features and probabilistic contextual aggregation features . Subsequently, the probabilistic context aggregates features It is input into the cross-scale attention fusion module for spatial feature scale calibration, feature pixel shuffling, nonlinear correction unit transformation, horizontal and vertical attention score calculation, as well as fusion, correction and channel compression transformation to perform target saliency reasoning and prediction; finally, the present invention adopts a saliency-aware collaborative optimization loss function to supervise the saliency prediction map output by the model, where the optimization target of the prediction map is guided by the saliency truth map (GT).

[0100] The proposed framework for adaptively camouflaged target recognition in complex scenarios includes an adaptive multi-scale feature enhancer, a probabilistic context aggregation module, a cross-scale attention fusion module, and a saliency-aware collaborative optimization loss function. The specific calculation methods and formulas for each part are as follows.

[0101] 1. Adaptive Multi-Scale Feature Enhancer

[0102] Step 1: Input features Perform convolution calculations and cross-kernel operations separately to obtain multi-branch output features with different receptive fields 、 、 and .

[0103] (1)

[0104] in Represents the multi-scale spatial features extracted by ResNet-50, Indicates that the convolution kernel size is The convolution operation, stands for cross-kernel operation.

[0105] Step 2: Feature concatenation and nonlinear correction units along the channel dimension are applied to the multi-branch output features 、 、 and , thereby generating channel weighted features .

[0106] (2)

[0107] in represents the feature concatenation operation in the channel dimension, Represents the nonlinear correction unit operation consisting of Conv, BN and ReLU.

[0108] Step 3: The channel squeeze and excitation module is combined with the channel multiplication operation and applied to the channel weighted features , thereby obtaining channel fusion features , is defined as follows:

[0109] (3)

[0110] in represents the channel squeeze and excitation module, and is a channel multiplication operation.

[0111] Step 4: Residual connection operation, element-wise addition operation and nonlinear correction unit are applied to the input features and channel fusion features , to generate visual enhancement features .

[0112] (4)

[0113] in Represents residual learning, + represents element-by-element addition operation, Represents the nonlinear correction unit operation.

[0114] 2. Cross-core operation

[0115] Step 1: The cross operation involves the input features Horizontal and vertical convolution operations are applied sequentially to generate features .

[0116] (5)

[0117] in and Represent the horizontal and vertical convolution operations respectively, and Represents the convolution kernel matrix.

[0118] Step 2: Generate features Perform matrix multiplication and softmax function normalization to obtain the generated features Horizontal attention score .

[0119] (6)

[0120] in represents matrix multiplication, represents the transpose transformation, represents the feature dimension, 、 and represents the weights used to generate query, keyword, and value vectors in the horizontal attention branch, Represents the softmax normalization function.

[0121] Step 3: Generate features Perform matrix multiplication and softmax function normalization to obtain the generated features Vertical attention score .

[0122] (7)

[0123] in represents matrix multiplication, represents the transpose transformation, represents the feature dimension, , and represents the weights used to generate query, keyword, and value vectors in the vertical attention branch, Represents the softmax normalization function.

[0124] Step 4: Score the horizontal attention and vertical attention score Use element-wise addition, Hadamard product, and nonlinear correction unit operations to generate branch output features , which is expressed as follows:

[0125] (8)

[0126] Where + represents element-by-element addition operation, represents the Hadamard product operation, represents the softmax normalization function, Indicates use Sub-nonlinear correction unit operation.

[0127] 3. Probabilistic Context Aggregation Module

[0128] Step 1: In the probability map feature Calculate the probability offset and generate probability predictions for foreground, background and confusion regions respectively 、 and .

[0129] (9)

[0130] in Represents the global threshold, the initial value is set to 0.5, represents metric normalization, represents the bilinear interpolation operation, represents the Sigmoid function, Represents convolution operation and channel compression transformation.

[0131] Step 2: In the local graph features Calculate the local probability offset and generate local foreground probability predictions respectively and local background probability prediction . They are represented as:

[0132] (10)

[0133] in Represents the local threshold, the initial value is set to 0.5, represents metric normalization, represents the bilinear interpolation operation, represents the Sigmoid function, Represents convolution operation and channel compression transformation.

[0134] Step 3: Global Probability Prediction 、 and and local probability predictions and Channel concatenation and nonlinear correction are performed to generate fused probability features is represented as follows:

[0135] (11)

[0136] in represents the feature concatenation operation in the channel dimension, Represents the nonlinear correction unit operation.

[0137] Step 4: Aggregate features and fusion probability features Perform scale reconciliation, feature reshaping, and probability prediction, and use Hadamard product to calculate context features .

[0138] (12)

[0139] in represents the bilinear interpolation operation, represents feature reshaping, Represents tensor rearrangement by dimension transformation, Represents convolution operation and channel compression transformation.

[0140] Step 5: Refine the salient objects by using similarity matrix calculation and value calculation to obtain attention scores and refinement features , defined as follows:

[0141] (13)

[0142] in represents matrix multiplication, represents the transpose transformation, represents the feature dimension, represents the softmax normalization function, , and Both represent convolution kernel matrices.

[0143] Step 6: Use spatial attention mechanism and nonlinear transformation to capture feature dependencies in spatial dimensions to obtain attention features as follows:

[0144] (14)

[0145] in represents the Hadamard product operation, represents the ReLU function, represents the Sigmoid function, and Represents the convolution kernel matrix.

[0146] Step 7: Aggregate features of input and attention features The fused and refined features are normalized by the nonlinear correction unit and the final output is The mathematical expression is:

[0147] (15)

[0148] in represents the feature cascade operation on the channel dimension, Indicates the use of cubic nonlinear correction unit operation.

[0149] 4. Cross-scale Attention Fusion Module

[0150] Step 1: Input features Use transposed convolution and pixel shuffling for upsampling, followed by multi-scale convolution operations to generate local features 、 and .

[0151] (16)

[0152] in represents an upsampling operation using transposed convolution and pixel shuffling, Indicates size Multi-scale convolution kernel .

[0153] Step 2: Local features 、 and Perform channel dimension feature cascade and refine the cascade features through the nonlinear correction unit to obtain refined features .

[0154] (17)

[0155] Step 3: Refine the features The horizontal attention score is generated by matrix multiplication and normalization function calculation and horizontally weighted features .

[0156] (18)

[0157] in , and represents the weights used to generate query, keyword, and value vectors in the horizontal attention branch, represents matrix multiplication, represents the transpose transformation, represents the feature dimension, Represents the softmax normalization function.

[0158] Step 4: Refine the features The vertical attention score is generated by matrix multiplication and normalization function calculation and vertical weighted features .

[0159] (19)

[0160] in , and represents the weights used to generate query, keyword, and value vectors in the vertical attention branch, represents matrix multiplication, represents the transpose transformation, represents the feature dimension, Represents the softmax normalization function.

[0161] Step 5: Horizontally weight features and vertical weighted features are fused, rectified, and channel-compressed to obtain the final output features as follows:

[0162] (20)

[0163] Where + represents element-by-element addition operation, Indicates the use of a quadratic nonlinear correction unit, Represents the channel compression transform.

[0164] As a result, the optimization strategy:

[0165] This paper proposes a saliency-aware collaborative optimization loss function, which consists of five weighted parts: (1) binary cross entropy loss ; (2) Joint intersection loss ; (3) Mean square error loss ; (4) Structural similarity index measurement ; (5) Mean absolute error loss These five components complement each other and can improve the accuracy of saliency prediction while refining the generated saliency prediction map. The saliency-aware collaborative optimization loss function is defined as follows:

[0166] First, stage The loss function is composed of binary cross entropy loss and joint intersection loss Together, they are defined as follows: this combination can effectively improve the model's robustness to noise interference, complex backgrounds, and category imbalance.

[0167] (twenty one)

[0168] Secondly, stage The loss function is achieved by combining the mean square error loss and structural similarity index metrics The combination can simultaneously optimize pixel-level accuracy and structural similarity, thereby improving the accuracy and visual quality of the prediction results.

[0169] (twenty two)

[0170] In addition, stage The loss function is based on the mean absolute error loss To build a more natural and accurate saliency map, we balance global and local information.

[0171] (twenty three)

[0172] Finally, saliency perception co-optimizes the loss function It can be expressed in weighted sum form as:

[0173] (twenty four)

[0174] The weight coefficient is set to , and .

[0175] As the technical effects of this embodiment:

[0176] In this embodiment, the model is trained on the DUTS-TR dataset with additional annotations, and the model is tested and evaluated on nine datasets including DUTS-TE, ECSSD, HKU-IS, PASCAL-S, DUT-OMRON, SCAS, DAVIS-S, HRSOD-TE and UHRSD-TE. The information of each training and test dataset is as follows: the DUTS-TR / TE dataset is used in both training and testing stages. It includes 10,553 training images and 5,019 test images, representing a comprehensive dataset for salient detection; the ECSSD dataset includes 1,000 test images with large-scale salient objects; the HKU-IS dataset consists of 4,447 complex images, most of which have features of medium size and multiple objects; the PASCAL-S dataset consists of 850 images, which are characterized by complex foreground objects and complex backgrounds; the DUT-OMRON dataset consists of 5,168 high-quality images, accompanied by various The SCAS dataset contains 2,554 test images, covering a variety of visual scenes, supporting research on target detection and multi-target recognition; the DAVIS-S dataset contains 92 high-resolution test images, provides pixel-level annotations, and covers salient targets in complex scenes; the HRSOD-TE dataset contains 400 high-resolution test images, covering a variety of scenes such as natural, artificial, and complex backgrounds, focusing on salient target detection in static images, challenging detail retention under high resolution; the UHRSD-TE dataset contains 988 high-resolution test images, including complex backgrounds, small targets, and multi-category salient objects.

[0177] Tables 1 and 2 show the comprehensive comparison results of the proposed method with existing advanced methods, including the performance of 16 methods such as C2S, RAS, and BASNet on six low-resolution benchmark datasets. The evaluation uses four widely recognized indicators: F β Metrics, mean absolute error MAE, E ϕ Metrics and S αThe experimental results show that in most test scenarios, the proposed method always shows better performance than other existing methods, verifying its robustness and effectiveness in low-resolution scenarios. It is particularly noteworthy that compared with the CoLA method, the proposed method S α The proposed method achieves significant improvement in metrics and significantly reduces MAE on multiple datasets. Specifically, on six low-resolution databases, the proposed method S α measure( F β The performance of the proposed method is significantly improved compared to the classic NKL method. S α measure( E ϕ The improvement of the metrics (accuracy, accuracy, and performance) is 1.4% (0.6%), 0.8% (-0.3%), 1.5% (0.3%), 2.7% (1.1%), 1.4% (0.6%), and -0.5% (5.2%), and the corresponding MAE is reduced by 2.9%, 3.1%, 0%, 6.0%, 8.9%, and 2.2%, respectively. These results highlight that the proposed method has strong accuracy and robustness in target prediction accuracy. Similarly, the proposed method also outperforms DCNet and achieves consistent progress on all evaluated datasets. F β measure( S α The performance of the proposed method is shown in Table 2. The performance of the proposed method is shown in Table 2. The performance of the proposed method is shown in Table 2. The performance of the proposed method is shown in Table 2. The performance of the proposed method is shown in Table 2. The performance of the proposed method is shown in Table 2. The performance of the proposed method is shown in Table 2. The performance of the proposed method is shown in Table 2. The performance of the proposed method is shown in Table 2. The performance of the proposed method is shown in Table 2. The performance of the proposed method is shown in Table 2. The performance of the proposed method is shown in Table 2. E ϕ measure(S α metric) by 1.7% (4.3%), 0.5% (2.3%), 1.2% (3.0%), 2.6% (4.9%), and 2.0% (4.3%), respectively, with corresponding MAE reductions of 20.6%, 12.5%, 7.1%, 10.0%, and 23.2%. While no single method universally dominates all metrics, the proposed method consistently ranks among the best, frequently surpassing state-of-the-art methods such as VST, ICON, and ISAANet in terms of overall effectiveness.

[0178] Table 3 shows a comprehensive quantitative comparison of the proposed method with the state-of-the-art methods on three high-resolution benchmark datasets. The evaluation uses four well-established metrics: F β Metrics, mean absolute error MAE, E ϕ Metrics and S α Metrics. Experimental results show that the proposed method can consistently provide excellent performance on almost all evaluation indicators of high-resolution datasets. The robust performance of multiple indicators strongly verifies the effectiveness and practicality of the proposed method in high-resolution scenarios. It is worth noting that in most cases, the proposed method shows significant advantages over 11 established techniques, including RAS, BASNet, U2Net, GCPANet, LDF, VST, ICON, DCNet, NKL, ISAANet and CoLA. Particularly noteworthy comparisons involve CoLA, where the proposed method achieves substantial improvements: in three datasets, S α measure( E ϕ The performance of the proposed method is significantly improved compared with the classic NKL method. On the three high-resolution databases, F β measure( S α The proposed method outperforms LDF, achieving consistent improvements on all evaluated datasets.S α The metrics increased by 2.3%, 2.8% and 3.9% respectively, while the MAE decreased by 26.7%, 6.7% and 80.8% respectively. These improvements represent a significant increase in prediction accuracy, and the reduction in MAE indicates a particularly significant error minimization. The consistent performance improvement across all evaluation metrics and datasets demonstrates that the proposed method successfully addresses key challenges in high-resolution image processing. Although the absolute performance varies slightly across different datasets, the proposed method maintains consistently high rankings in all evaluations. This stability across a variety of high-resolution scenarios further highlights the robustness of the proposed method and its potential for real-world applications where consistent performance is critical. The combination of these quantitative results demonstrates that the proposed method not only achieves state-of-the-art progress in raw performance metrics, but also provides a reliable and practical solution for high-resolution image analysis tasks.

[0179] In addition, an evaluation was conducted in this example to determine the significant performance of the method proposed in the present invention. As shown in Table 4, the indicators used were average F β Metrics (avg- F β ), average joint intersection (avg-miou) and mean absolute error (avg-mae). Experimental results show that compared with the latest CoLA (PiNet), the method proposed in this paper reduces avg- F β , avg-miou and avg-mae indicators are improved by 4.4% (0.5%), 14.8% (6.6%) and 55.0% (15.0%). In addition, the proposed method also shows better performance improvements compared with classic methods such as RAS, BASNet, GCPANet, VST and NKL. Specifically, the proposed method improves the avg-miou (avg-mae) indicators by 4.2% (62.5%), 5.1% (25.0%), 2.3% (10.0%), 0.9% (10.0%) and 2.6% (5.0%), respectively. In summary, the proposed method has been shown to achieve a higher level of performance indicators for the salient object detection task. Although its detection performance fails to optimize all indicators for all datasets, it is also very close to these maximum performance levels, which is a good indication of its effectiveness and reliability.

[0180] This example also uses some challenging scenes to qualitatively demonstrate the visual results of the proposed method and other state-of-the-art methods in low-resolution and high-resolution saliency maps, including large objects, multiple objects, simple scenes, complex scenes, and low-contrast scenes. Figure 2As shown in the figure, it is clear that the proposed method can more accurately identify, localize, and segment salient object details and boundaries, resulting in salient predictions closer to the ground truth in many cases. Specifically, rows 1 and 3 demonstrate that the proposed method achieves excellent performance in segmenting large objects in simple scenes. For example, the image shows an offshore multi-manipulator oil production equipment (row 1). Most existing methods cannot fully accurately detect, localize, and segment the manipulators and base, while the proposed method can clearly identify fine objects and reduce segmentation errors. Similarly, in row 3, there is a wooden pavilion located on green grass. The proposed method can more clearly delineate the pavilion's hollow structure, demonstrating that the proposed method is more competitive in processing the details of large objects. Rows 4 and 5 demonstrate that the proposed method has good ability to identify and predict dual objects in simple scenes. For example, in row 4, there are two athletes running vigorously on a track. Compared with other methods, the proposed method can more accurately identify and segment the morphological information of the two athletes, and the saliency prediction is very close to the ground truth. Furthermore, the proposed method also performs well in identifying salient objects in complex scenes (rows 6 and 7). Specifically, there are two swimmers waiting to disembark in a swimming pool (row 6). The proposed method can more clearly locate and predict the swimmers' overall shape and arm posture than other existing methods, without the missed detections and false positives common to other methods. Similarly, as shown in rows 2 and 10, the proposed method also demonstrates excellent performance in single-object recognition and segmentation in low-contrast scenes. For example, in row 2, there is an egret standing on the shore. The proposed method predicts the egret's overall shape, but relatively lacks information about its head and beak. Compared to other existing methods, the proposed method does not produce false positives. For multi-object recognition and prediction in low-contrast scenes, such as rows 9, 11, and 12, the proposed method can describe the details and shapes of multiple objects in greater detail without missing any objects, and is more competitive with state-of-the-art methods such as CoLA and PiNet. For example, in row 12, there are five sparrows foraging on hay. The proposed method accurately predicts all sparrows, with target locations highly consistent with the ground-truth map. Compared to other existing methods, the proposed method does not miss any detections. In summary, quantitative and qualitative comparisons fully demonstrate the effectiveness and reliability of the proposed method in the task of salient object detection and recognition.

[0181] In order to more comprehensively evaluate the network architecture, the flexibility and efficiency of the proposed method and other existing methods were evaluated, such as RAS, SCRN, LDF and other 15 methods. Table 5 provides the architecture size, parameter count and FLOPs of various existing methods. All data measurements were performed on a server using an RTX 4070Ti GPU, and the image size was 384×384. As shown in Table 5, the proposed method has 28.25M parameters, 70.63G FLOPs and a model size of 107.9MB. At the same time, there are obvious differences between the proposed method and its counterparts, such as U2Net, GCPANet, GateNet, EDRNet, SHNet, EGNet and MINst. Significant changes in the number of parameters (FLOPs) are observed, with the proposed method changing by -55.9% (-20.3%), -137.4% (-10.8%), -355.2% (-59.4%), -39.2% (-34.6%), +36.4% (-56.6%), -295.4% (-245.6%) and -23.2% (-474.9%), respectively. Comparison of the proposed method with the classic heavyweight methods RAS, SCRN, LDF, ICON, PGNet, NKL, ECF-DT, and VST shows that the model size (number of parameters) has changed significantly, which are +14.2% (+14.4%), +11.6% (+12.0%), +12.5% ​​(+12.3%), -16.8% (-17.0%), -156.9% (-157.2%), +10.8% (+10.6%), +0.1% (+0.5%), and -65.3% (-56.1%), respectively. In summary, the proposed method is an effective method for salient object detection. Although it is not the lightest in all three indicators, its performance is relatively outstanding, which is acceptable.

[0182] Table 1 Quantitative comparison of the proposed method and the state-of-the-art methods on three low-resolution benchmark databases

[0183] ,

[0184] Table 2 Quantitative comparison of the proposed method and the state-of-the-art methods on three low-resolution benchmark databases

[0185] ,

[0186] Table 3 Quantitative comparison of the proposed method and the state-of-the-art methods on three high-resolution benchmark databases

[0187] ,

[0188] Table 4 Quantitative comparison of the method of the present invention and the existing method provided in the embodiment of the present invention

[0189] ,

[0190] Table 5 Comparison and analysis of the flexibility of various methods

[0191] ,

[0192] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0193] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A method for identifying adaptive camouflaged targets in complex scenarios, characterized by: include: Acquire an image to be tested; the image to be tested is a captured image with a complex scene; Inputting the image to be tested into the target recognition model for processing to obtain a target recognition result; The object recognition model is constructed based on an interactive guidance network and a saliency-aware collaborative optimization loss function; the interactive guidance network includes an adaptive multi-scale feature enhancer, a probabilistic context aggregation module, and a cross-scale attention fusion module connected in sequence; the saliency-aware collaborative optimization loss function consists of five weighted components: a binary cross-entropy loss, a joint intersection loss, a mean square error loss, a structural similarity index metric, and a mean absolute error loss; The image to be tested is input into the target recognition model for processing to obtain the target recognition result. The specific process includes: The ResNet-50 feature encoder is used to extract multi-scale spatial features of the image to be tested, and multi-scale features from low-level features to high-level features are obtained. 、 、 、 and Among them, the characteristics and For low-level features, features 、 and For advanced features; The multi-scale features 、 、 、 and Input the adaptive multi-scale feature enhancer to enhance visual information and generate visual enhancement features 、 、 、 and ; For the multi-scale features 、 、 、 and And the visual enhancement features 、 、 、 and Perform feature concatenation and channel compression transformation to generate probability map features , local graph features and aggregate features , and input the generated features into the probability context aggregation module to calculate the probability context aggregation features ; Aggregate the probabilistic context features Input the cross-scale attention fusion module for prediction to obtain the final target recognition result; The operation process of the adaptive multi-scale feature enhancer specifically includes: Step 1: Input features Perform convolution calculations and cross-kernel operations separately to obtain multi-branch output features with different receptive fields 、 、 and : ， in, Represents the multi-scale spatial features extracted by ResNet-50, Indicates that the convolution kernel size is The convolution operation, stands for cross-kernel operation; Step 2: Feature concatenation and nonlinear correction units along the channel dimension are applied to the multi-branch output features 、 、 and , thereby generating channel weighted features : ， in, represents the feature concatenation operation in the channel dimension, Represents the nonlinear correction unit operation composed of Conv, BN and ReLU; Step 3: Combine the channel squeeze and excitation module with the channel multiplication operation and apply it to the channel weighted features , thereby obtaining channel fusion features : ， in, represents the channel squeezing and excitation module, Represents channel multiplication operation; Step 4: Residual connection operation, element-wise addition operation and nonlinear correction unit are applied to the input features and channel fusion features , to generate visual enhancement features : ， in, Represents residual learning, + represents element-by-element addition operation, Represents the nonlinear correction unit operation; The formula of the saliency-aware collaborative optimization loss function is expressed as: ， in, Co-optimize loss function for saliency perception; 、 and is the weight coefficient, set to , and ; For the Binary cross entropy loss at the stage Combined loss sum; For the The mean squared error loss of the stage Structural similarity index sum; For the The mean absolute error loss of the stage .

2. The method for identifying adaptive camouflaged targets in complex scenarios according to claim 1, characterized in that: The cross-core operation process specifically includes: Step 1: Using the cross operation involves the input features Horizontal and vertical convolution operations are applied sequentially to generate features : ， in, and denote horizontal and vertical convolution operations respectively, and Represents the convolution kernel matrix; Step 2: Generate features Perform matrix multiplication and softmax function normalization to obtain the generated features Horizontal attention score : ， in, represents matrix multiplication, represents the transpose transformation, represents the feature dimension, 、 and represents the weights used to generate query, keyword, and value vectors in the horizontal attention branch, represents the softmax normalization function; Step 3: Generate features Perform matrix multiplication and softmax function normalization to obtain the generated features Vertical attention score : ， in, represents matrix multiplication, represents the transpose transformation, represents the feature dimension, , and represents the weights used to generate query, keyword, and value vectors in the vertical attention branch, represents the softmax normalization function; Step 4: Score the horizontal attention and vertical attention score Use element-wise addition, Hadamard product, and nonlinear correction unit operations to generate branch output features , which is expressed as follows: ， Among them, + represents element-by-element addition operation, represents the Hadamard product operation, represents the softmax normalization function, Indicates use Sub-nonlinear correction unit operation.

3. The method for identifying adaptive camouflaged targets in complex scenarios according to claim 1, characterized in that: The computation process of the probabilistic context aggregation module specifically includes: Step 1: In the probability map feature Calculate the probability offset and generate probability predictions for foreground, background and confusion regions respectively 、 and : ， in, Represents the global threshold, the initial value is set to 0.5, represents metric normalization, represents the bilinear interpolation operation, represents the Sigmoid function, Represents convolution operation and channel compression transformation; Step 2: In the local graph features Calculate the local probability offset and generate local foreground probability predictions respectively and local background probability prediction , respectively expressed as: ， in, Represents the local threshold, the initial value is set to 0.5, represents metric normalization, represents the bilinear interpolation operation, represents the Sigmoid function, Represents convolution operation and channel compression transformation; Step 3: Global Probability Prediction 、 and and local probability predictions and Channel concatenation and nonlinear correction are performed to generate fused probability features , expressed as: ， in, represents the feature concatenation operation in the channel dimension, Represents the nonlinear correction unit operation; Step 4: Aggregate features and fusion probability features Perform scale reconciliation, feature reshaping, and probability prediction, and use Hadamard product to calculate context features : ， in, represents the bilinear interpolation operation, represents feature reshaping, Represents tensor rearrangement by dimension transformation, Represents convolution operation and channel compression transformation; Step 5: Refine the salient objects by using similarity matrix calculation and value calculation to obtain attention scores and refinement features : ， in, represents matrix multiplication, represents the transpose transformation, represents the feature dimension, represents the softmax normalization function, , and Both represent convolution kernel matrices; Step 6: Use spatial attention mechanism and nonlinear transformation to capture feature dependencies in spatial dimensions to obtain attention features as follows: ， in, represents the Hadamard product operation, represents the ReLU function, represents the Sigmoid function, and Represents the convolution kernel matrix; Step 7: Aggregate features of input and attention features The fused and refined features are normalized by the nonlinear correction unit and the final output is The mathematical expression is: ， in, represents the feature cascade operation on the channel dimension, Indicates the use of cubic nonlinear correction unit operation.

4. The method for identifying adaptive camouflaged targets in complex scenarios according to claim 1, characterized in that: The operation process of the cross-scale attention fusion module specifically includes: Step 1: Input features Use transposed convolution and pixel shuffling for upsampling, followed by multi-scale convolution operations to generate local features 、 and : ， in represents an upsampling operation using transposed convolution and pixel shuffling, Indicates size Multi-scale convolution kernel ; Step 2: Local features 、 and Perform channel dimension feature cascade and refine the cascade features through the nonlinear correction unit to obtain refined features : ， Step 3: Refine the features The horizontal attention score is generated by matrix multiplication and normalization function calculation and horizontally weighted features : ， in, , and represents the weights used to generate query, keyword, and value vectors in the horizontal attention branch, represents matrix multiplication, represents the transpose transformation, represents the feature dimension, represents the softmax normalization function; Step 4: Refine the features The vertical attention score is generated by matrix multiplication and normalization function calculation and vertical weighted features : ， in , and represents the weights used to generate query, keyword, and value vectors in the vertical attention branch, represents matrix multiplication, represents the transpose transformation, represents the feature dimension, represents the softmax normalization function; Step 5: Horizontally weight features and vertical weighted features are fused, rectified, and channel-compressed to obtain the final output features as follows: ， Among them, + represents element-by-element addition operation, Indicates the use of a quadratic nonlinear correction unit, Represents channel compression transform.

5. A recognition system for adaptively camouflaged targets in complex scenarios, applying the method according to any one of claims 1 to 4, characterized in that: include: An image acquisition unit, configured to acquire an image to be tested; the image to be tested is a captured image with a complex scene; The target recognition unit is used to input the image to be tested into the target recognition model for processing to obtain a target recognition result; the target recognition model is constructed based on an interactive guidance network and a saliency-aware collaborative optimization loss function; the interactive guidance network includes an adaptive multi-scale feature enhancer, a probabilistic context aggregation module, and a cross-scale attention fusion module connected in sequence; the saliency-aware collaborative optimization loss function consists of five weighted parts, namely binary cross entropy loss, joint intersection loss, mean square error loss, structural similarity index measurement, and mean absolute error loss.

6. An electronic device, characterized in that: The electronic device comprises a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the method for identifying adaptive camouflaged targets in complex scenes according to any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that The device stores a computer program, which, when executed by a processor, implements the method for identifying adaptive camouflaged targets in complex scenes as claimed in any one of claims 1 to 4.

Images