A target recognition method, system, device and medium for complex scenes

Abstract

The present invention discloses a target recognition method, system, device and medium for complex scenes, and relates to the field of computer vision technology. The method comprises: inputting a test image into a target recognition model for processing to obtain a target recognition result; the target recognition model is constructed based on a dual-path coding network and a weighted optimization loss function; the dual-path coding network includes a MobileNet path and a wavelet path; wherein the MobileNet path is used to extract multi-scale spatial features and is composed of a hierarchical encoder and a multi-feature prediction decoder connected in sequence; the wavelet path is used to extract frequency domain features and is composed of a global semantic wavelet coding module and a local wavelet fusion decoding module connected in sequence; the multi-feature prediction decoder is also connected to the global semantic wavelet coding module and the local wavelet fusion decoding module, respectively. The present invention can improve the precision and accuracy of target recognition for complex scenes.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to a target recognition method, system, device and medium for complex scenes. Background Art

[0002] In the context of salient object recognition and prediction, the core goal is to automatically identify and extract the most discriminative salient regions or objects from complex images or videos. For example, in criminal science applications, these salient objects may be key evidence at a crime scene (such as a murder weapon or bloodstains), suspicious individuals or vehicles in surveillance videos, or abnormal regions in forensic images. Traditional saliency detection methods primarily rely on low-level image features (such as color, texture, and edges), but they have certain limitations when dealing with complex scenes, including the following issues:

[0003] 1. Limited detection performance in complex scenes: Salient object detection performance often degrades when dealing with complex scenes (e.g., multiple objects, cluttered backgrounds, or changing lighting). Traditional methods rely on low-level features, struggle to distinguish objects from backgrounds, and can result in false or missed detections in extremely complex scenes.

[0004] 2. Inadequate detection of small objects and edge details: Existing methods perform poorly when processing small objects or object edge details. Because small objects occupy fewer pixels in an image, their features are easily overwhelmed by background information. Loss of object edge details leads to inaccurate saliency maps, hindering subsequent analysis and application.

[0005] 3. High computing resource requirements: Although deep learning-based methods can capture high-level semantic information, they typically require a large amount of computing resources and storage space, especially when processing high-resolution images or videos. This makes it difficult for these methods to run efficiently on devices with limited computing power (such as mobile devices or embedded systems), limiting their popularity in practical applications.

[0006] 4. Limited Generalization: Existing salient object detection models have weak generalization capabilities and struggle to adapt to new scenarios or unknown objects. Most models are trained on specific datasets, which can overfit the distribution of the training data, leading to significant performance degradation when applied across domains or scenarios. Furthermore, a lack of diverse training data also limits the model's generalization capabilities. Summary of the Invention

[0007] The purpose of the present invention is to provide a target recognition method, system, device and medium for complex scenes, which can improve the precision and accuracy of target recognition for complex scenes.

[0008] To achieve the above object, the present invention provides the following solutions:

[0009] A target recognition method for complex scenes, comprising:

[0010] Acquire an image to be tested; the image to be tested is a captured image with a complex scene;

[0011] 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 a dual-path coding network and a weighted optimization loss function; the dual-path coding network includes a MobileNet path and a wavelet path; wherein, the MobileNet path is used to extract multi-scale spatial features, and is composed of a hierarchical encoder and a multi-feature prediction decoder connected in sequence; the wavelet path is used to extract frequency domain features, and is composed of a global semantic wavelet coding module and a local wavelet fusion decoding module connected in sequence; the multi-feature prediction decoder is also connected to the global semantic wavelet coding module and the local wavelet fusion decoding module, respectively.

[0012] Optionally, the weighted optimization loss function includes binary cross entropy loss, intersection-over-union loss and mean square error loss.

[0013] Optionally, the operation process of the global semantic wavelet coding module specifically includes:

[0014] Step 1: By using the discrete wavelet transform based on Haar wavelet, the input feature map obtained from the image to be tested is transformed into Decomposed into four sub-band graphs , the subband graph Used as low-frequency component to represent the input The basic characteristics of the subband graph , subband diagram sum subband diagram Used as high frequency component to represent the output detail:

[0015] ,

[0016] in, Represents discrete wavelet transform based on Haar wavelet;

[0017] Step 2: Low frequency components and high frequency components 、 and Perform depth-separable convolution and nonlinear activation operations to obtain features 、 、 and Respectively expressed as:

[0018] ,

[0019] in, Indicates that the convolution kernel size is Depthwise separable convolution operation, Represents the GeLU function;

[0020] Step 3: Apply nonlinear correction unit and normalization function to the features 、 、 and , to obtain the feature weights 、 、 as well as :

[0021] ,

[0022] in, represents the average pooling operation, Represents the nonlinear correction unit operation composed of DSConv, BN and ReLU, Represents the Sigmoid function;

[0023] Step 4: Low frequency features and high-frequency features 、 and Perform feature weighting to obtain weighted features 、 、 and :

[0024] ,

[0025] in, represents the Hadamard product operation;

[0026] Step 5: Multi-feature fusion is performed by using element-wise addition operation, which is expressed as:

[0027] ,

[0028] in, Represents the nonlinear correction unit operation composed of DSConv, BN and GeLU;

[0029] Step 6: Obtain downsampled features by using a nonlinear correction unit operation and , the final output feature It is expressed as follows:

[0030] ,

[0031] in, represents the channel dimension feature cascade, Indicates use The nonlinear correction unit operation consists of DSConv, BN and GeLU.

[0032] Optionally, the operation process of the local wavelet fusion decoding module specifically includes:

[0033] Step 1: Input features Perform resizing and nonlinear correction operations to obtain high-resolution output features :

[0034] ,

[0035] in, represents the bilinear interpolation algorithm, Indicates the use of two nonlinear correction unit operations consisting of DSConv, BN and ReLU;

[0036] Step 2: Features Perform wavelet decomposition to obtain the low-frequency subband map and high frequency subband diagram 、 and , respectively expressed as:

[0037] ,

[0038] in, Represents discrete wavelet transform based on Haar wavelet;

[0039] Step 3: Features 、 、 、 and features 、 、 、 Perform multi-feature fusion and transformation to obtain output features 、 、 and , respectively expressed as:

[0040] ,

[0041] Among them, + represents element-by-element addition operation;

[0042] Step 4: Features 、 、 and Perform inverse wavelet transform to obtain output features 、 、 and , respectively expressed as:

[0043] ,

[0044] in, represents the inverse wavelet transform;

[0045] Step 5: Multi-region features are processed by using feature concatenation and convolution operations 、 、 and Reshape and fuse, and finally output features It is expressed as follows:

[0046] ,

[0047] in, and Represents the corresponding channel dimension feature cascade, Indicates use The nonlinear correction unit operation consists of DSConv, BN and GeLU.

[0048] Optionally, the operation process of the multi-feature prediction decoder specifically includes:

[0049] Step 1: Compress high-level semantic information through nonlinear correction unit and adaptive average pooling to obtain compressed semantic kernel and It is expressed as follows:

[0050] ,

[0051] in, represents adaptive average pooling, represents the (DSConv+ReLU) operation, and Represent the fifth-stage features encoded by global semantic wavelet and MobileNet respectively;

[0052] Step 2: Use cross attention and element-wise addition operations to perform spatial attention calibration and obtain attention features and It is expressed as follows:

[0053] ,

[0054] in, represents the crisscross attention mechanism, 、 、 and Represent the third-stage and fourth-stage features encoded by global semantic wavelet coding and MobileNet, respectively;

[0055] Step 3: Multi-scale hierarchical feature fusion is performed on the compressed semantic kernel 、 and attention features 、 Thus, the directional enhancement of low-level features is achieved through high-level semantics, and the enhanced features are expressed as:

[0056] ,

[0057] in, and have different expansion ratios Compressed semantic kernel and ;

[0058] Step 4: Use residual links and nonlinear transformations to improve shallow multi-channel features 、 、 and After fusion and optimization, the features after fusion and optimization are expressed as:

[0059] ,

[0060] in, Represents an operation consisting of DSConv, BN, and Sigmoid, represents the PReLU function, 、 、 and Represent the first-stage and second-stage features encoded by global semantic wavelet coding and MobileNet, respectively;

[0061] Step 5: Channel attention calibration is performed by recalibrating the channel feature responses using feature reshaping, tensor multiplication, and normalization transformation to amplify the semantic attribute discrimination and output features 、 、 and It is expressed as follows:

[0062] ,

[0063] in, represents feature reshaping, represents the Softmax function, represents the matrix multiplication operation, represents the transposed matrix;

[0064] Step 6: Iteratively integrate spatial features by using lightweight convolutional blocks, specifically expressed as:

[0065] ,

[0066] in, Represents the feature concatenation of the channel dimension;

[0067] Step 7: Aggregate by using skip connections to generate the final output , which is expressed as follows:

[0068] ,

[0069] in, Indicated by DSConv , Up and DCSonv operations, It is The wavelet decoding characteristics of the stage, Indicates the Stage wavelet features The output features after performing the scale adjustment operation and the nonlinear correction unit operation, and Respectively represent Global semantic wavelet coding and MobileNet encoding features at the stage.

[0070] The present invention also provides a target recognition system for complex scenes, comprising:

[0071] 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;

[0072] 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 a dual-path coding network and a weighted optimization loss function; the dual-path coding network includes a MobileNet path and a wavelet path; wherein, the MobileNet path is used to extract multi-scale spatial features and is composed of a hierarchical encoder and a multi-feature prediction decoder connected in sequence; the wavelet path is used to extract frequency domain features and is composed of a global semantic wavelet coding module and a local wavelet fusion decoding module connected in sequence; the multi-feature prediction decoder is also connected to the global semantic wavelet coding module and the local wavelet fusion decoding module respectively.

[0073] 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 target recognition method for complex scenarios.

[0074] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the complex scene-oriented target recognition method as described above.

[0075] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0076] The present invention discloses a method, system, device, and medium for target recognition in complex scenes. The method comprises inputting a target image into a target recognition model for processing to obtain a target recognition result. The target recognition model is constructed based on a dual-path coding network and a weighted optimization loss function. The dual-path coding network includes a MobileNet path and a wavelet path. The MobileNet path is used to extract multi-scale spatial features and is composed of a hierarchical encoder and a multi-feature prediction decoder connected in sequence. The wavelet path is used to extract frequency domain features and is composed of a global semantic wavelet encoding module and a local wavelet fusion decoding module connected in sequence. The multi-feature prediction decoder is also connected to the global semantic wavelet encoding module and the local wavelet fusion decoding module, respectively. The present invention can improve the precision and accuracy of target recognition in complex scenes. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] 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.

[0078] Figure 1 This is a flow chart of the target recognition method for complex scenarios provided in this embodiment;

[0079] Figure 2 A comparison chart of Avg-Fmeasure, Avg-mIOU, and Avg-MAE of the method of the present invention and other comparative methods provided in this embodiment;

[0080] Figure 3 This is a comparison chart of the significance prediction results using the method of the present invention and other methods provided in this embodiment. DETAILED DESCRIPTION

[0081] 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.

[0082] The purpose of the present invention is to provide a target recognition method, system, device and medium for complex scenes, which can improve the precision and accuracy of target recognition for complex scenes.

[0083] 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.

[0084] like Figure 1 As shown, the present invention provides a target recognition method for complex scenes, comprising:

[0085] Step 100: Acquire an image to be tested; the image to be tested is a captured image with a complex scene.

[0086] Step 200: Input the 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 a dual-path coding network and a weighted optimization loss function; the dual-path coding network includes a MobileNet path and a wavelet path; wherein the MobileNet path is used to extract multi-scale spatial features and is composed of a hierarchical encoder and a multi-feature prediction decoder connected in sequence; the wavelet path is used to extract frequency domain features and is composed of a global semantic wavelet coding module and a local wavelet fusion decoding module connected in sequence; the multi-feature prediction decoder is also connected to the global semantic wavelet coding module and the local wavelet fusion decoding module, respectively.

[0087] Based on the above technical solution, the following specific implementation methods are provided.

[0088] (1) System Introduction

[0089] This paper proposes a lightweight salient object recognition and prediction method for complex scenes. Figure 1 It is worth noting that the method of the present invention uses MobileNet as the encoding path branch to extract multi-scale spatial features, which are respectively denoted as 、 、 、 and At the same time, the coding path branch based on the global semantic wavelet coding module is also used to extract frequency domain features, which is recorded as: 、 、 、 and .from Figure 1 It can be seen that and Represent low-level features to high-level features, respectively, 、 、 and It has rich positioning information and target details, which is more conducive to obtaining salient boundaries; as a high-level feature, 、 、 、 、 and Contains abstract semantic information, which is more conducive to understanding the target category. First, use high-level features and Generate convolution kernels with different dilation ratios for the multi-feature prediction decoder. As the input of the local wavelet fusion decoding module to obtain richer frequency domain features 、 、 and ; Subsequently, multi-path features 、 and It is input into the multi-feature prediction decoder for spatial attention calibration, semantically guided feature enhancement, feature reshaping, iterative refinement of lightweight convolution blocks, and skip connection fusion to perform reasoning and prediction of salient targets; finally, the present invention uses a weighted optimization loss function to comprehensively supervise the output saliency prediction map of the proposed method, which is supervised by the saliency ground truth map GT.

[0090] The lightweight salient object recognition and prediction framework for complex scenarios proposed in this scheme includes a global semantic wavelet encoding module, a local wavelet fusion decoding module, a multi-feature prediction decoder, and a weighted optimization loss function. The specific calculation methods and formulas for each part are as follows.

[0091] 1. Global semantic wavelet coding module

[0092] Step 1: By using the discrete wavelet transform based on Haar wavelet, the input feature map Decomposed into four sub-band graphs. Sub-band graph Used as low-frequency component to represent the input The basic characteristics of the subband diagram 、 and Used as high frequency component to represent the output detail.

[0093] (1)

[0094] in, Represents the discrete wavelet transform based on Haar wavelet.

[0095] Step 2: Low frequency components and high frequency components 、 and Perform depth-wise separable convolution and nonlinear activation operations to further learn and extract useful information, features 、 、 and Respectively expressed as:

[0096] (2)

[0097] in, Indicates that the convolution kernel size is Depthwise separable convolution operation, Represents the GeLU function.

[0098] Step 3: Apply nonlinear correction unit and normalization function to the features 、 、 and , to obtain the feature weights 、 、 as well as .

[0099] (3)

[0100] in, represents the average pooling operation, Represents the nonlinear correction unit (DSConv+BN+ReLU) operation, Represents the Sigmoid function.

[0101] Step 4: Low frequency features and high-frequency features 、 and Perform feature weighting to obtain weighted features 、 、 and .

[0102] (4)

[0103] in Represents the Hadamard product operation.

[0104] Step 5: Multi-feature fusion is performed by using element-wise addition operation, which makes the feature map have richer detail information. They are expressed as:

[0105] (5)

[0106] in Represents the (DSConv+ReLU) operation.

[0107] Step 6: and The integration is achieved by using a nonlinear correction unit, where low-frequency and high-frequency features work together to create a downsampled feature map that retains the full high-frequency and low-frequency information. The final output feature It is expressed as follows:

[0108] (6)

[0109] in represents the channel dimension feature cascade, Indicates use (DSConv+BN+GeLU) operations.

[0110] 2. Local wavelet fusion decoding module

[0111] Step 1: Input features Perform resizing and nonlinear correction operations to obtain high-resolution output features .

[0112] (7)

[0113] in represents the bilinear interpolation algorithm, Indicates the use of a secondary nonlinear correction unit operation.

[0114] Step 2: Features Perform wavelet decomposition to obtain the low-frequency subband map and high frequency subband diagram 、 and , which are respectively expressed as:

[0115] (8)

[0116] in Represents the discrete wavelet transform based on Haar wavelet.

[0117] Step 3: Features 、 、 、 and features 、 、 、 Perform multi-feature fusion and transformation to obtain output features 、 、 and .

[0118] (9)

[0119] Where + represents element-by-element addition.

[0120] Step 4: Features 、 、 and Perform inverse wavelet transform to obtain output features 、 、 and .

[0121] (10)

[0122] in Represents the inverse wavelet transform.

[0123] Step 5: Multi-region features are processed by using feature concatenation and convolution operations 、 、 and Reshape and fuse, and finally output features It is expressed as follows:

[0124] (11)

[0125] 3. Multi-feature prediction decoder

[0126] Step 1: Compress high-level semantic information through nonlinear correction unit and adaptive average pooling to obtain compressed semantic kernel and It is expressed as follows:

[0127] (12)

[0128] in represents adaptive average pooling, represents the (DSConv+ReLU) operation, and Represent the fifth-stage features encoded by global semantic wavelet coding and MobileNet, respectively.

[0129] Step 2: Use cross attention and element-wise addition operations to perform spatial attention calibration to reduce multi-feature spatial misalignment. and It is expressed as follows:

[0130] (13)

[0131] in represents the crisscross attention mechanism, 、 、 and Represent the third-stage and fourth-stage features encoded by global semantic wavelet coding and MobileNet, respectively.

[0132] Step 3: Multi-scale hierarchical feature fusion is performed on the compressed semantic kernel 、 and attention features 、 Thus, the targeted enhancement of low-level features is achieved through high-level semantics.

[0133] (14)

[0134] in and have different expansion ratios Compressed semantic kernel and .

[0135] Step 4: Use residual links and nonlinear transformations to improve shallow multi-channel features 、 、 and Fusion and optimization are performed to improve feature representation.

[0136] (15)

[0137] in Represents the (DSConv+BN+Sigmoid) operation, represents the PReLU function, 、 、 and denote the first-stage and second-stage features encoded by global semantic wavelet encoding and MobileNet, respectively.

[0138] Step 5: Channel attention calibration is performed by recalibrating the channel feature responses using feature reshaping, tensor multiplication, and normalization transformation to amplify the semantic attribute discrimination. Output features 、 、 and It is expressed as follows:

[0139] (16)

[0140] in represents feature reshaping, represents the Softmax function, represents the matrix multiplication operation, Represents the transposed matrix.

[0141] Step 6: Progressive fusion refinement of features is achieved by iteratively integrating spatial features using lightweight convolutional blocks, thereby suppressing noise and improving boundary accuracy.

[0142] (17)

[0143] in Represents the feature concatenation of the channel dimension.

[0144] Step 7: The consistency between global semantics and local details is ensured by using skip connection aggregation to refine the features to generate the final output. The mathematical formula is as follows:

[0145] (18)

[0146] in Representation (DSConv +Up+DCSonv) operation, It is The wavelet decoding characteristics of the stage, Indicates the Stage wavelet features The output features after performing the scale adjustment operation and the nonlinear correction unit operation, and Respectively represent Global semantic wavelet coding and MobileNet encoding features at the stage.

[0147] (2) Result optimization strategy

[0148] This paper proposes a weighted optimization loss function that can remove interference information and noise in the image while enhancing the saliency reasoning and prediction capabilities. The loss function consists of three parts: (1) Binary cross entropy loss Able to calculate the loss of each pixel; (2) Intersection-over-Union loss Able to optimize the global structure of the input image; (3) Mean square error loss It can further enhance the details of salient objects. These three loss calculation parts complement each other to improve the accuracy of saliency prediction. 、 and In the The stages are defined as follows:

[0149] (19)

[0150] (20)

[0151] (twenty one)

[0152] in , It's a pixel The predicted probability of It's a pixel The true value label of and Both are calculated by the multi-feature prediction decoder.

[0153] Weighted optimization loss function As shown in mathematical formula (22):

[0154] (twenty two)

[0155] in and .

[0156] As a specific application example, this example trains the model on the DUTS-TR dataset with additional annotations, and tests and evaluates the model on the DUTS-TE, ECSSD, HKU-IS, PASCAL-S, and DUT-OMRON datasets. Information about the training and testing datasets is as follows: The DUTS-TR / TE dataset, used in both training and testing phases, consists of 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 are medium-sized and feature multiple objects; the PASCAL-S dataset consists of 850 images, characterized by complex foreground objects and complex backgrounds; and the DUT-OMRON dataset consists of 5,168 high-quality images with a wide variety of object categories and complex scenes.

[0157] In this embodiment, the F-measure and MAE numerical comparison of the method of the present invention and other existing cutting-edge methods on five public datasets, DUTS-TE, ECSSD, HKU-IS, PASCAL-S, and DUT-OMRON, are shown in Table 1. As can be seen from the results shown in Table 1, the method of the present invention is always superior to other existing lightweight cutting-edge methods, such as CSNet, SAMNet, FSMINet, CorrNet, and BDFGNet, in most cases. This finding helps to verify the effectiveness and reliability of the method of the present invention. It is particularly noteworthy that when MAE (F-measure) is used on the five public datasets, the method of the present invention achieves 0.043 (0.805), 0.043 (0.81), 0.034 (0.892), 0.063 (0.786), and 0.055 (0.753) on the five datasets, respectively. Specifically, our method demonstrates superior performance compared to the classic CSNet, with F-measure (MAE) scores on five public datasets improving by 17.2% (72.1%), 4.4% (51.2%), 6.2% (73.5%), 8.7% (63.5%), and 11.6% (47.3%), respectively. Compared to CorrNet, our method improves F-measure (MAE) values ​​on the five datasets by 2.7% (39.5%), 0.8% (74.4%), 2.5% (52.9%), 1.2% (68.3%), and 3.3% (21.8%). Furthermore, when using F-measure (MAE) values ​​on the DUTS-TE, HKU-IS, and DUT-OMRON datasets, our method outperforms the state-of-the-art BDFGNet by 1.4% (14.0%), 1.1% (23.5%), and 4.1% (5.5%), respectively. Although our method does not achieve the highest numerical results on the ECSSD dataset, it achieves relatively optimal performance compared to other similar methods (such as CorrNet, EDN, and BDFGNet), demonstrating the good performance of our method. As expected, there are some performance differences under certain metrics, which are considered reasonable and acceptable. Overall, our method outperforms other mainstream lightweight methods across all datasets using these evaluation metrics.

[0158] In addition, this example also uses Avg-Fmeasure, Avg-mIOU and Avg-MAE indicators to evaluate the significance performance of the method of the present invention. Figure 2The results show that compared with FSMINet (HVPNet), the proposed method improves the average performance of F-measure, MAE and mIOU metrics by 6.0% (2.5%), 41.7% (29.2%) and 8.8% (2.8%) respectively. In general, for lightweight saliency prediction methods, the proposed method has achieved a high level in various detection performance indicators. Although its detection performance is not as good as that of heavyweight methods, it is approaching the performance level of these methods. Figure 3 It can also be clearly seen from the saliency prediction visual results in that the method of the present invention is powerful enough in visual prediction performance and can meet the detection performance requirements of most visual tasks.

[0159] Table 1 Comparison of F-measure and MAE values ​​of the proposed method and other comparison methods on five data sets

[0160] .

[0161] This example also provides a comparison of the visual results of target saliency prediction using the method of the present invention and other state-of-the-art methods in five public datasets using some challenging scenes. Figure 3 As shown in , it contains some challenging scenes, such as simple scenes, complex scenes, large objects, multiple objects, and low contrast. Figure 2 The following conclusions can be drawn: Our method accurately detects and delineates object boundaries closer to the ground truth across a range of scenarios, demonstrating high accuracy and efficiency. Specifically, our method achieves superior performance in large object prediction and segmentation, as demonstrated by the salient object predictions observed in rows 2, 14, and 15. For example, the image in row 2 shows a middle-aged man walking across a street accompanied by a white dog. Clearly, the saliency map predicted by our method is closer to the ground truth, accurately predicting the dog's morphology and tail features, surpassing many leading lightweight methods, including ADMNet, CorrNet, and CSNet. As demonstrated in rows 5, 8, 9, and 16, the effectiveness of our method lies in its ability to accurately identify and segment multiple salient objects in complex scenes. Clearly, our method is able to accurately predict object boundaries or describe individual morphologies in more complex scenes, including a scene of multiple people pitching a tent (row 16). This superior performance is a significant advantage compared to existing leading-edge methods. In addition, the proposed method also shows excellent performance in scenes with low contrast between the target and the image background, e.g. Figure 3As shown in the 1st, 4th, 11th and 13th rows of the dataset, the proposed method can more clearly depict the shape and position of the target, achieving better performance than other existing lightweight saliency detection methods. For example, the 4th row depicts an egret standing in the middle of the water with the blue river behind it. The proposed method clearly depicts the details of the egret's head and legs, outperforming FSMINet, HVPNet and SAMNet. It is worth noting that the performance of the proposed method is comparable to that of heavyweight methods such as MSRMNet, OLER and VST. Overall, compared with other lightweight methods, the proposed method has demonstrated excellent performance in target localization, prediction and segmentation on five public datasets.

[0162] Therefore, this scheme proposes a lightweight salient target recognition and prediction method for complex scenes. In terms of technical settings, considering the superiority of frequency domain features for target saliency recognition, a dual-path encoding network is first designed as the baseline for saliency reasoning, which includes a MobileNet path to obtain spatial domain features and a wavelet path based on a global semantic wavelet coding module to capture frequency domain features. In order to effectively encode the input image and generate multi-level features, thereby achieving adaptive edge enhancement with low computational complexity, a global semantic wavelet coding module is proposed and integrated into the dual-path encoding network. The local wavelet fusion decoding module dynamically integrates high-frequency detail information (such as edges and textures) with low-frequency semantic features through frequency-aware fusion, achieving accurate reconstruction of fine-grained structures without sacrificing computational efficiency. The multi-feature prediction decoder uses cross-attention to coordinate heterogeneous feature representations extracted from the hierarchical encoder layers, aiming to resolve semantic ambiguity between multi-resolution features, and improves prediction consistency through attention-guided fusion, ensuring computational efficiency and enhanced interpretability. In terms of optimization settings, this paper proposes a weighted optimization loss function that enhances saliency inference and prediction capabilities while removing interfering information and noise from the image. This loss function consists of three components: a binary cross-entropy loss that calculates the loss for each pixel; an intersection-over-union loss that optimizes the global structure of the input image; and a mean squared error loss that further enhances salient object details. These three losses complement each other to improve the accuracy of saliency prediction. Experimental results on five public datasets demonstrate that this method outperforms existing cutting-edge methods.

[0163] 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.

[0164] 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 target recognition method for complex scenes, characterized in that: 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 target recognition model is constructed based on a dual-path encoding network and a weighted optimization loss function; the dual-path encoding network includes a MobileNet path and a wavelet path; wherein the MobileNet path is used to extract multi-scale spatial features and is composed of a hierarchical encoder and a multi-feature prediction decoder connected in sequence; the wavelet path is used to extract frequency domain features and is composed of a global semantic wavelet encoding module and a local wavelet fusion decoding module connected in sequence; the multi-feature prediction decoder is also connected to the global semantic wavelet encoding module and the local wavelet fusion decoding module respectively; The operation process of the global semantic wavelet coding module specifically includes: Step 1: By using the discrete wavelet transform based on Haar wavelet, the input feature map obtained from the image to be tested is transformed into Decomposed into four sub-band graphs , the subband graph Used as low-frequency component to represent the input The basic characteristics of the subband graph , subband diagram sum subband diagram Used as high frequency component to represent the output detail: ， in, Represents discrete wavelet transform based on Haar wavelet; Step 2: Low frequency components and high frequency components 、 and Perform depth-separable convolution and nonlinear activation operations to obtain features 、 、 and Respectively expressed as: ， in, Indicates that the convolution kernel size is Depthwise separable convolution operation, Represents the GeLU function; Step 3: Apply nonlinear correction unit and normalization function to the features 、 、 and , to obtain the feature weights 、 、 as well as : ， in, represents the average pooling operation, Represents the nonlinear correction unit operation composed of DSConv, BN and ReLU, Represents the Sigmoid function; Step 4: Low frequency features and high-frequency features 、 and Perform feature weighting to obtain weighted features 、 、 and : ， in, represents the Hadamard product operation; Step 5: Multi-feature fusion is performed by using element-wise addition operation, which is expressed as: ， in, Represents the nonlinear correction unit operation composed of DSConv, BN and GeLU; Step 6: Obtain downsampled features by using a nonlinear correction unit operation and , the final output feature It is expressed as follows: ， in, represents the channel dimension feature cascade, Indicates use The nonlinear correction unit operation consists of DSConv, BN and GeLU; The operation process of the local wavelet fusion decoding module specifically includes: Step 1: Input features Perform resizing and nonlinear correction operations to obtain high-resolution output features : ， in, represents the bilinear interpolation algorithm, Indicates the use of two nonlinear correction unit operations consisting of DSConv, BN and ReLU; Step 2: Features Perform wavelet decomposition to obtain the low-frequency subband map and high frequency subband diagram 、 and , respectively expressed as: ， in, Represents discrete wavelet transform based on Haar wavelet; Step 3: Features 、 、 、 and features 、 、 、 Perform multi-feature fusion and transformation to obtain output features 、 、 and , respectively expressed as: ， Among them, + represents element-by-element addition operation; Step 4: Features 、 、 and Perform inverse wavelet transform to obtain output features 、 、 and , respectively expressed as: ， in, represents the inverse wavelet transform; Step 5: Multi-region features are processed by using feature concatenation and convolution operations 、 、 and Reshape and fuse, and finally output features It is expressed as follows: ， in, and Represents the corresponding channel dimension feature cascade, Indicates use The nonlinear correction unit operation consists of DSConv, BN and GeLU.

2. The target recognition method for complex scenes according to claim 1, characterized in that: The weighted optimization loss functions include binary cross entropy loss, intersection-over-union loss, and mean square error loss.

3. The target recognition method for complex scenes according to claim 1, characterized in that: The operation process of the multi-feature prediction decoder specifically includes: Step 1: Compress high-level semantic information through nonlinear correction unit and adaptive average pooling to obtain compressed semantic kernel and It is expressed as follows: ， in, represents adaptive average pooling, represents the (DSConv+ReLU) operation, and Represent the fifth-stage features encoded by global semantic wavelet and MobileNet respectively; Step 2: Use cross attention and element-wise addition operations to perform spatial attention calibration and obtain attention features and It is expressed as follows: ， in, represents the crisscross attention mechanism, 、 、 and Represent the third-stage and fourth-stage features encoded by global semantic wavelet coding and MobileNet, respectively; Step 3: Multi-scale hierarchical feature fusion is performed on the compressed semantic kernel 、 and attention features 、 Thus, the directional enhancement of low-level features is achieved through high-level semantics, and the enhanced features are expressed as: ， in, and have different expansion ratios Compressed semantic kernel and ; Step 4: Use residual links and nonlinear transformations to improve shallow multi-channel features 、 、 and After fusion and optimization, the features after fusion and optimization are expressed as: ， in, Represents an operation consisting of DSConv, BN, and Sigmoid, represents the PReLU function, 、 、 and Represent the first-stage and second-stage features encoded by global semantic wavelet coding and MobileNet, respectively; Step 5: Channel attention calibration is performed by recalibrating the channel feature responses using feature reshaping, tensor multiplication, and normalization transformation to amplify the semantic attribute discrimination and output features 、 、 and It is expressed as follows: ， in, represents feature reshaping, represents the Softmax function, represents the matrix multiplication operation, represents the transposed matrix; Step 6: Iteratively integrate spatial features by using lightweight convolutional blocks, specifically expressed as: ， in, Represents the feature concatenation of the channel dimension; Step 7: Aggregate by using skip connections to generate the final output , which is expressed as follows: ， in, Indicated by DSConv , Up and DCSonv operations, It is The wavelet decoding characteristics of the stage, Indicates the Stage wavelet features The output features after performing the scale adjustment operation and the nonlinear correction unit operation, and Respectively represent Global semantic wavelet coding and MobileNet encoding features at the stage.

4. A target recognition system for complex scenes, applying the method according to any one of claims 1 to 3, 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 a dual-path coding network and a weighted optimization loss function; the dual-path coding network includes a MobileNet path and a wavelet path; wherein, the MobileNet path is used to extract multi-scale spatial features and is composed of a hierarchical encoder and a multi-feature prediction decoder connected in sequence; the wavelet path is used to extract frequency domain features and is composed of a global semantic wavelet coding module and a local wavelet fusion decoding module connected in sequence; the multi-feature prediction decoder is also connected to the global semantic wavelet coding module and the local wavelet fusion decoding module respectively.

5. 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 target recognition method for complex scenes according to any one of claims 1 to 3.

6. A computer-readable storage medium, characterized in that It stores a computer program, which, when executed by a processor, implements the target recognition method for complex scenes as described in any one of claims 1 to 3.

Images