Target detection and identification method based on self-adaptive search network guidance

Abstract

The invention discloses a target detection and identification method based on self-adaptive search network guidance, and relates to the field of target detection. According to the method, a multi-layer feature interaction and fusion structure is established, a search module based on adaptive attention is designed to extract multi-feature (low, high, texture, edge and the like) information of a target, a multi-scale feature fusion structure is utilized to further retain high-precision multi-scale fusion features, the positioning recognition precision based on different semantics and space information is improved, and the target positioning recognition precision is improved. And the target detection recognition rate is improved.

Description

Technical Field

[0001] The present invention relates to the field of target detection, and in particular to a target detection and recognition method based on adaptive search network guidance, which is used to solve the problems of reduced accuracy of general target detection methods due to weak automatic detection capability of target features, fuzzy target position information, poor adaptive robustness, etc. in complex scenes. Background Art

[0002] The statements in this section merely provide background information related to the present disclosure and may not constitute prior art.

[0003] Object detection is one of the basic tasks in the field of computer vision. It is widely used in real life and work scenarios, such as intelligent transportation systems (vehicle recognition, pedestrian detection, etc.), security monitoring (automatic identification of abnormal behavior, such as intrusion detection, etc.), retail (customer behavior analysis, inventory management, etc.), medical health (auxiliary diagnosis, such as tumor detection, disease screening, etc.), drone technology (for environmental monitoring, crop management, search and rescue missions, etc.). The progress of object detection has promoted the development of the entire computer vision field. With the rise of deep learning, object detection algorithms based on convolutional neural networks (CNNs) have achieved remarkable results, which not only improves the accuracy of detection, but also speeds up the detection speed, making real-time applications possible. From the R-CNN and OverFeat proposed in 2013, to the later Fast / FasterR-CNN, SSD, YOLO series, and the recently emerging Transformer architecture. In terms of network structure, many good algorithm technologies have emerged in the object detection technology based on deep learning, from two stage to one stage, from bottom-up only to Top-Down, from single scale network to feature pyramid network, and from PC-oriented to mobile-oriented. These algorithms have excellent detection effects and performance on open object detection datasets.

[0004] These methods can be roughly divided into three types. One is to design targeted network modules / architectures to effectively explore the characteristics of the target area. One is to incorporate some auxiliary tasks into the joint learning / multi-task learning framework, such as classification tasks, edge extraction, salient target detection, etc. This method can mine valuable additional clues from shared features and significantly enhance the feature representation of the target. Another is the bionic method, which simulates the behavioral process of predators in nature or human visual psychological patterns to design networks to better identify targets. Zhao et al. (2019) proposed a salient target detection method based on edge-guided network (EGNet). Zhang et al. (2022) proposed a multi-scale full-scene monitoring target detection method based on attention mechanism. Wu et al. (2023) proposed an adaptive feature enhanced target detection network (YOLO-AFENet) to improve the detection accuracy of small targets. Liu et al. (2022) proposed a multi-scale salient target detection network method integrating attention mechanism. Wang et al. (2023) proposed a salient target detection network based on semantic information guided feature aggregation.

[0005] Although the recently proposed target detection and recognition methods have made significant progress, there are still some problems. For example, due to the destruction of the edge of the target area or the blurred outline, there are shortcomings such as weak feature extraction network performance, insufficient image semantic information, and inability to automatically search for target location features, resulting in low target recognition accuracy and low object detection robustness. In addition, the above methods are basically feature extraction operators and network structures designed according to specific scenarios. After training, they cannot be adaptively migrated and applied to target detection tasks in new scenarios, affecting the promotion and application of detection and recognition algorithms. Summary of the invention

[0006] The purpose of the present invention is to provide a target detection and recognition method based on adaptive search network guidance to solve the problems existing in the above-mentioned prior art. By building a multi-layer feature interaction and fusion structure, a search module based on adaptive attention is designed to extract multi-feature (low, high, texture, edge, etc.) information of the target, and a multi-scale feature fusion structure is used to further retain high-precision multi-scale fusion features, thereby improving the positioning and recognition accuracy based on different semantics and spatial information, and improving the target detection recognition rate, thereby solving the above-mentioned problems.

[0007] The technical solution of the present invention is as follows:

[0008] A target detection and recognition method based on adaptive search network guidance comprises: based on the Yolov8 detector framework, introducing an adaptive attention feature screening module, a search network SANet guided learning based on a dual-channel receptive field MRF module, and a multi-scale feature fusion module to form a new detection network; and performing target detection and recognition on images acquired by a sensor through the detection network.

[0009] Furthermore, the adaptive attention feature screening module guides the network to adaptively select useful target features under different receptive fields from the dimensions of channels and space; the dual-channel receptive field MRF module is used to expand the receptive field to capture more and more discriminative features, and use dense connections to save feature information from different levels; the multi-scale feature fusion module classifies the target category and regresses the position through the fusion feature sequence of multiple scales and cross-layer mixed connections, and integrates multi-level fusion features to predict the target object.

[0010] Furthermore, the adaptive attention feature screening module includes: an input layer, a convolution layer connected to the input layer, a channel self-attention module, a spatial self-attention module and an output layer; the convolution layer includes: three parallel convolutions; the outputs of two convolutions are sent to the channel self-attention module for feature screening in the channel domain, and fused and outputted, and are simultaneously input into the spatial self-attention module with the output of the other convolution for feature screening in the spatial domain, and finally outputted by the output layer.

[0011] Furthermore, the three parallel convolutions are: 3*3 convolution, 5*5 convolution and 3*3 dilated convolution; the outputs of 5*5 convolution and 3*3 dilated convolution are sent to the channel self-attention module for channel domain feature screening and fusion output. F output by another 3*3 convolution S1 At the same time, it is input into the spatial self-attention module for feature screening in the spatial domain, and finally output by the output layer F out .

[0012] Furthermore, the dilation rate of the dilated convolution is 3.

[0013] Furthermore, the receptive field MRF module includes: five branches; in each branch, the size of the first convolutional layer is 1×1, which is used to reduce the number of channels to 32;

[0014] The two layers after the first branch are: a 3×3 convolutional layer composed of a 1×3 convolutional layer and a 3×1 convolutional layer, and a 3×3 atrous convolutional layer with a dilation rate of 3;

[0015] The two layers after the second branch are: a 5×5 convolutional layer composed of a 1×5 convolutional layer and a 5×1 convolutional layer, and a 3×3 dilated convolutional layer with a dilation rate of 5;

[0016] The last two layers of the third branch are: a 7×7 convolutional layer composed of a 1×7 convolutional layer and a 7×1 convolutional layer, and a 3×3 atrous convolutional layer with a dilation rate of 7;

[0017] After the first four branches are connected in parallel, the number of channels is reduced to 32 through a 1×1 convolution operation, and they are input into the ReLU function as a whole together with the fifth branch to obtain features.

[0018] Furthermore, the multi-scale feature fusion module uses channel domain maximum pooling and channel domain average pooling operations to input the two generated compressed features into the same MLP module, and then concatenates the output features, multiplies the result obtained after Sigmoid function processing with the input feature to obtain a channel attention map.

[0019] Furthermore, the specific operation process of the multi-scale feature fusion module is shown in the following formula:

[0020] M c (F) = σ(M LP (P Avg (F)); M LP (P Max (F)))

[0021] in:

[0022] M c (F) indicates output;

[0023] F is the input feature;

[0024] σ is the Sigmoid function;

[0025] M LP Represents the MLP module, which is a multi-layer perceptron consisting of a convolutional layer and a ReLU activation function;

[0026] P Avg With P Max They represent channel domain average pooling and channel domain maximum pooling respectively;

[0027] ; Indicates a concatenation operation.

[0028] Furthermore, target detection and recognition is performed on the image acquired by the sensor through the detection network, including:

[0029] Step S1: sending the image acquired by the sensor to the detection network;

[0030] Step S2: The first feature extraction module captures the target feature points, and then the adaptive attention feature screening module selects the target features of different importance, deeply mines the low-level, high-level, texture, semantic and other feature information of the target's intrinsic relevance and collects them to obtain the mapping feature vector

[0031] Step S3: Performed simultaneously with step S2, the search network SANet based on the dual-channel receptive field MRF module actively searches for the target area, guides the detection network to learn features, and obtains the mapping feature vector

[0032] Step S4: Map feature vector and the mapping feature vector Perform point multiplication to obtain the weighted feature vector

[0033] Step S5: The features outputted from step S4 are then passed through the second feature extraction module to capture the target feature points, and the adaptive attention feature screening module is used to screen out target features of different importance levels to obtain the feature vector

[0034] Step S6: Performed simultaneously with step S5, the search network SANet based on the dual-channel receptive field MRF module actively searches for the target area, guides the detection network to learn features, and obtains the mapping feature vector

[0035] Step S7: Map feature vector and the mapping feature vector Perform point multiplication to obtain the weighted feature vector

[0036] Step S8: The features outputted from step S7 are then passed through the third feature extraction module to capture the target feature points, and the adaptive attention feature screening module is used to screen out target features of different importance levels, so as to further capture richer feature information vectors.

[0037] Step S9: Performed simultaneously with step S7, the search network SANet based on the dual-channel receptive field MRF module actively searches for the target area, guides the detection network to learn features, and obtains a new mapping feature vector

[0038] Step S10: Map feature vector and the mapping feature vector Perform point multiplication to obtain the weighted feature vector

[0039] Step S11: The features outputted from step S4, step S7 and step S10 are respectively subjected to a channel attention operation composed of maximum pooling and average pooling, and the two generated compressed features are inputted into the same MLP module, and then the output features are concatenated, and the result obtained after Sigmoid function processing is multiplied by the input feature to obtain three channel attention maps of different scales;

[0040] Step S12: further fuse the multi-scale features obtained in step S11 to obtain the final comprehensive detection and recognition result.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] 1. Aiming at target detection in complex task scenarios, the present invention introduces an adaptive attention module to achieve feature screening of different types of target images, making the network's learning of target features more effective;

[0043] 2. Aiming at the weak automatic detection capability of target features and fuzzy target location information of existing methods, the present invention proposes a search network SANet based on a dual-channel receptive field MRF module. The search network actively searches for target areas and performs feature learning, thereby improving the high-quality characterization capability of target features and the detection performance.

[0044] 3. The present invention utilizes a multi-scale feature fusion module and integrates multi-level fusion features to enhance the target object prediction capability and further improve the detection performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a diagram of the target detection and recognition network framework based on the guidance of the adaptive search network;

[0046] Figure 2 The structural diagram of the adaptive attention feature screening module;

[0047] Figure 3 This is a diagram of the search network structure based on the dual-channel receptive field MRF module;

[0048] Figure 4 This is the structure diagram of the multi-scale feature fusion module. DETAILED DESCRIPTION

[0049] It should be noted that relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0050] The features and performance of the present invention are further described in detail below in conjunction with the embodiments.

[0051] Embodiment 1

[0052] See also Figure 1 , a target detection and recognition method based on adaptive search network guidance, comprising:

[0053] Based on the Yolov8 detector framework, an adaptive attention feature screening module, a search network SANet guided learning based on a dual-channel receptive field MRF module, and a multi-scale feature fusion module are introduced to construct a new detection network; the detection network is used to perform target detection and recognition on images acquired by the sensor.

[0054] In this embodiment, specifically, the adaptive attention feature screening module guides the network to adaptively select useful target features under different receptive fields from the dimensions of channels and spaces, thereby broadening the comprehensiveness of information;

[0055] The dual-channel receptive field MRF module is used to expand the receptive field to capture more and more distinguishing features, and uses dense connections to save feature information from different levels, thereby enhancing the feature quality and attention level of the target area, thereby improving the detection performance; that is, based on the Yolov8 detector framework, feature screening of the adaptive attention mechanism is introduced, and the search network SANet guided learning based on the dual-channel receptive field MRF module is realized, so that the spatial and contextual semantic information of the image target can be fully utilized, the multi-scale feature information of the image target can be adaptively learned, and the feature weights can be automatically adjusted according to the image itself, so as to adapt to target detection in different scenarios; that is, the search network SANet of the dual-channel receptive field MRF module actively searches for the target area and learns the target features;

[0056] The multi-scale feature fusion module classifies the target category and regresses the position through the fusion feature sequence of multiple scales and cross-layer hybrid connection, integrates multi-level fusion features to predict the target object, and further improves the detection accuracy; that is, the two pooling methods of the channel attention mechanism are used to make the network pay more attention to effective information, such as texture information and salient areas with low color contrast;

[0057] It should be noted that the target detection and recognition network proposed in this embodiment is based on the improved YOLOv8 framework. By making full use of the target position and contextual semantics actively extracted by the search network, the backbone network is guided to enhance the learning of the target feature area, and the multi-scale attention feature fusion structure is used to perform feature fusion and recognition of the target at different levels; the network model can adaptively learn the multi-scale features of the target area according to the target image of the task scene, and adaptively adjust the feature weights according to the image itself, and perform multi-scale detection and recognition, thereby improving the recognition effect;

[0058] In this embodiment, specifically, the adaptive attention mechanism is unique in that it can adaptively screen target features of different importance, and an adaptive attention feature screening module is introduced into the backbone network. Low-level, high-level, texture, contextual semantics and other multi-information extraction is performed through training to capture richer features. By deeply mining and collecting the feature information of the intrinsic correlation of the target, rich and reliable information basis is provided for subsequent target detection tasks. At the same time, the adaptive attention feature screening module can guide the network to adaptively select useful target features under different receptive fields from the dimensions of channels and space, broadening the comprehensiveness of information and improving fault tolerance.

[0059] For details, please refer to Figure 2 , the adaptive attention feature screening module comprises:

[0060] An input layer, a convolutional layer connected to the input layer, a channel self-attention module, a spatial self-attention module and an output layer; the convolutional layer includes: three parallel convolutions to obtain feature maps with different receptive fields; the three parallel convolutions are: 3*3 convolution, 5*5 convolution and 3*3 dilated convolution, and the dilation rate of the dilated convolution is 3; the feature maps obtained by the three convolutional layers are respectively expressed as:

[0061] F 3 =W 3×3 ×F input

[0062] F 5 =W 5×5 ×F input

[0063]

[0064] Among them, the two outputs of 5*5 convolution and 3*3 dilated convolution are sent to the channel self-attention module for channel domain feature screening and fusion output. F output by another 3*3 convolution S1 At the same time, it is input into the spatial self-attention module for feature screening in the spatial domain, and finally output by the output layer F out ;

[0065] That is, by aggregating information from multiple cores, the size of the neuron's adaptive receptive field can be achieved. The adaptive attention mechanism is unique in that it can adaptively screen target features of different importance levels, and low-level, high-level, texture, contextual semantics and other information are extracted through training to capture richer feature information. The feature information of the intrinsic correlation of the target is deeply mined and collected to provide rich and reliable information basis for subsequent target detection tasks. At the same time, this module can guide the network to adaptively select useful target features under different receptive fields from the dimensions of channels and space, broadening the comprehensiveness of information and improving the effective performance of the detector.

[0066] In this embodiment, specifically, a receptive field MRF module is used to expand the receptive field and obtain more distinguishing features in the search stage, dense connections are used to save information from different levels, and the target area is actively found through the search network SANet based on the dual-channel receptive field MRF module, thereby guiding the backbone network to learn the target features;

[0067] That is, the perception ability of the receptive field module MRF for tiny spatial displacements is used and embedded into each stage of the backbone network. The target position information captured at each stage is used to continuously guide and correct the backbone network's learning of the target feature area, thereby capturing the target feature area in the entire image, narrowing the feature learning scope, and capturing feature information from different levels through dense connections, expanding the receptive field of the detection network, and thus improving the model's representation ability and performance.

[0068] For details, please refer to Figure 3 The receptive field MRF module (Multiple Receptive Field) includes five branches. In each branch, the size of the first convolutional layer is 1×1, which is used to reduce the number of channels to 32.

[0069] The two layers after the first branch are: a 3×3 convolutional layer composed of a 1×3 convolutional layer and a 3×1 convolutional layer, and a 3×3 atrous convolutional layer with a dilation rate of 3;

[0070] The two layers after the second branch are: a 5×5 convolutional layer composed of a 1×5 convolutional layer and a 5×1 convolutional layer, and a 3×3 dilated convolutional layer with a dilation rate of 5;

[0071] The last two layers of the third branch are: a 7×7 convolutional layer composed of a 1×7 convolutional layer and a 7×1 convolutional layer, and a 3×3 atrous convolutional layer with a dilation rate of 7;

[0072] After the first four branches are connected in parallel, the number of channels is reduced to 32 through a 1×1 convolution operation, and they are input into the ReLU function as a whole together with the fifth branch to obtain features.

[0073] In this embodiment, specifically, by designing convolution operations of different sizes, features of different scales are extracted in the same network layer. Subsequently, multiple parallel branches are used to extract features of different scales, and a channel attention mechanism is applied to the feature map of each scale to highlight information that is more meaningful to the current task.

[0074] Specifically, Figure 4 As shown, the multi-scale feature fusion module uses channel domain maximum pooling and channel domain average pooling operations to input the two generated compressed features into the same MLP module, and then splices the output features. The result obtained after Sigmoid function processing is multiplied by the input feature to obtain the channel attention map. The specific operation process is shown in the following formula:

[0075] M c (F) = σ(M LP (P Avg (F)); M LP (P Max (F)))

[0076] in:

[0077] M c (F) indicates output;

[0078] F is the input feature;

[0079] σ is the Sigmoid function;

[0080] M LP Represents the MLP module, which is a multi-layer perceptron consisting of a convolutional layer and a ReLU activation function;

[0081] P Avg With P Max They represent channel domain average pooling and channel domain maximum pooling respectively;

[0082] ; Indicates the splicing operation;

[0083] Finally, the multi-scale feature maps processed by the channel attention mechanism are fused and sent to the subsequent network layer for the final detection task decision;

[0084] That is, the multi-scale feature fusion module extracts features of different scales in the same network layer through convolution operations of different sizes, uses multiple parallel branches to extract features of different scales, and applies the channel attention mechanism to the feature map of each scale to highlight the information that is more meaningful to the current task. At the same time, the maximum pooling and average pooling operations are used to input the two generated compressed features into the same MLP module, and then the output features are spliced. The result obtained after Sigmoid function processing is multiplied by the input feature to obtain the channel attention map. Finally, the multi-scale feature maps processed by the channel attention mechanism are fused, and the fused feature maps are sent to the subsequent network layers, which enhances the decision-making ability of the detection task and improves the performance of the detector.

[0085] Embodiment 2

[0086] This embodiment is a specific application of the target detection and recognition method proposed in Embodiment 1, and its details are as follows:

[0087] (1) Network construction and reasoning:

[0088] The backbone network adopts an improved Darknet53-like structure, which specifically includes modules such as residual and spatial pooling pyramid. Then, the target features are fused together in a multi-layer feature interactive connection based on the dual-channel receptive field MRF module. Among them, the three convolution layers in the adaptive attention feature screening module are 3*3 convolution, 5*5 convolution and 3*3 dilated convolution, and the dilation rate of the dilated convolution is 3; two paths are sent to the channel self-attention module for feature screening in the channel domain, and fused and output; and the other path is simultaneously input into the spatial self-attention module for feature screening in the spatial domain.

[0089] After constructing a search network SANet based on a dual-channel receptive field MRF module, it is embedded into the backbone network to guide learning, which makes it possible to make full use of the spatial, contextual semantics and other information of the image target, adaptively learn the multi-scale feature information of the image target, and fuse the extracted features and re-input them into the backbone of another source to form a feature interactive fusion structure.

[0090] The connection Neck adopts an improved U-shaped recalculation bidirectional jump connection pyramid structure. The difference between it and the original recalculation pyramid structure is that the original U-shaped feature transfer loop is expanded to N U-shaped loops, and the feature is changed from layer-by-layer connection to a sequential and skip-level mixed connection mode, so that multi-scale features can be better integrated. This implementation is N=3.

[0091] The labeled target image is input into the target detection and recognition network model guided by the adaptive search network for training. During the training process, the commonly used training parameter settings in YOLOv8 are adopted, combined with the post-processing strategy. The trained model is saved, and the test image is input to test it to obtain the final target comprehensive detection and recognition results.

[0092] (2) Dataset preparation: In addition to using commonly used public target detection datasets, you can also build data for verification based on specific task scenarios. If the amount of data is insufficient, you can use data augmentation to expand it.

[0093] The target detection and recognition network guided by the adaptive search network makes full use of the spatial and contextual semantics of the image target, adaptively and actively learns the multi-scale features of the image target, automatically adjusts the feature weights according to the image itself, and improves the detection and recognition accuracy. This model can solve the problems of reduced detection accuracy due to the weak ability to automatically detect target features, fuzzy target location information, poor adaptive robustness, and the weakening of target features caused by the inability of the trained model to adapt to input images of various qualities.

[0094] In this embodiment, based on the detection network constructed above, target detection and recognition are performed on the image acquired by the sensor, which specifically includes the following steps:

[0095] Step S1: sending the image acquired by the sensor to the detection network;

[0096] Step S2: The first feature extraction module captures the target feature points, and then the adaptive attention feature screening module selects the target features of different importance, deeply mines the low-level, high-level, texture, semantic and other feature information of the target's intrinsic relevance and collects them to obtain the mapping feature vector

[0097] Step S3: Performed simultaneously with step S2, the search network SANet based on the dual-channel receptive field MRF module actively searches for the target area, guides the detection network to learn features, and obtains the mapping feature vector

[0098] Step S4: Map feature vectors and the mapping feature vector Perform point multiplication to obtain the weighted feature vector

[0099] Step S5: The features outputted from step S4 are then passed through the second feature extraction module to capture the target feature points, and the adaptive attention feature screening module is used to screen out target features of different importance levels to obtain the feature vector

[0100] Step S6: Performed simultaneously with step S5, the search network SANet based on the dual-channel receptive field MRF module actively searches for the target area, guides the detection network to learn features, and obtains the mapping feature vector

[0101] Step S7: Map feature vector and the mapping feature vector Perform point multiplication to obtain the weighted feature vector

[0102] Step S8: The features outputted from step S7 are then passed through the third feature extraction module to capture the target feature points, and the adaptive attention feature screening module is used to screen out target features of different importance levels, so as to further capture richer feature information vectors.

[0103] Step S9: Performed simultaneously with step S7, the search network SANet based on the dual-channel receptive field MRF module actively searches for the target area, guides the detection network to learn features, and obtains a new mapping feature vector

[0104] Step S10: Map feature vector and the mapping feature vector Perform point multiplication to obtain the weighted feature vector

[0105] Step S11: The features outputted from step S4, step S7 and step S10 are respectively subjected to a channel attention operation composed of maximum pooling and average pooling, and the two generated compressed features are inputted into the same MLP module, and then the output features are concatenated, and the result obtained after Sigmoid function processing is multiplied by the input feature to obtain three channel attention maps of different scales;

[0106] Step S12: further fuse the multi-scale features obtained in step S11 to obtain the final comprehensive detection and recognition result.

[0107] The above-mentioned embodiments only express the specific implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the protection scope of the present application. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the technical solution concept of the present application, and these all belong to the protection scope of the present application.

[0108] This background section is provided to generally present the context of the invention, and the work of the presently named inventors, the work to the extent described in this background section, and aspects of this section that did not constitute prior art at the time of application are neither explicitly nor implicitly admitted to be prior art to the present invention.

Claims

1. A target detection and recognition method based on adaptive search network guidance, characterized in that: include: Based on the Yolov8 detector framework, an adaptive attention feature screening module, a search network SANet guided learning based on a dual-channel receptive field MRF module, and a multi-scale feature fusion module are introduced to construct a new detection network; the detection network is used to perform target detection and recognition on images acquired by the sensor.

2. According to claim 1, a target detection and recognition method based on adaptive search network guidance is characterized in that: The adaptive attention feature screening module guides the network to adaptively select useful target features under different receptive fields from the dimensions of channels and space; the dual-channel receptive field MRF module is used to expand the receptive field to capture more and more discriminative features, and use dense connections to save feature information from different levels; the multi-scale feature fusion module classifies the target category and regresses the position through the fusion feature sequence of multiple scales and cross-layer mixed connections, and integrates multi-level fusion features to predict the target object.

3. The target detection and recognition method based on adaptive search network guidance according to claim 2 is characterized in that: The adaptive attention feature screening module includes: an input layer, a convolution layer connected to the input layer, a channel self-attention module, a spatial self-attention module and an output layer; the convolution layer includes: three parallel convolutions; the outputs of two convolutions are sent to the channel self-attention module for feature screening in the channel domain, and fused and outputted, and are simultaneously input into the spatial self-attention module with the output of the other convolution for feature screening in the spatial domain, and finally outputted by the output layer.

4. The target detection and recognition method based on adaptive search network guidance according to claim 3 is characterized in that: The three parallel convolutions are: 3*3 convolution, 5*5 convolution and 3*3 dilated convolution; the outputs of 5*5 convolution and 3*3 dilated convolution are sent to the channel self-attention module for channel domain feature screening and fusion output. F output by another 3*3 convolution S1 At the same time, it is input into the spatial self-attention module for feature screening in the spatial domain, and finally output by the output layer F out .

5. The target detection and recognition method based on adaptive search network guidance according to claim 4 is characterized in that: The dilation rate of the dilated convolution is 3.

6. The target detection and recognition method based on adaptive search network guidance according to claim 1, characterized in that: The dual-channel receptive field MRF module includes: five branches; in each branch, the size of the first convolution layer is 1×1, which is used to reduce the number of channels to 32; The two layers after the first branch are: a 3×3 convolutional layer composed of a 1×3 convolutional layer and a 3×1 convolutional layer, and a 3×3 atrous convolutional layer with a dilation rate of 3; The two layers after the second branch are: a 5×5 convolutional layer composed of a 1×5 convolutional layer and a 5×1 convolutional layer, and a 3×3 dilated convolutional layer with a dilation rate of 5; The last two layers of the third branch are: a 7×7 convolutional layer composed of a 1×7 convolutional layer and a 7×1 convolutional layer, and a 3×3 atrous convolutional layer with a dilation rate of 7; After the first four branches are connected in parallel, the number of channels is reduced to 32 through a 1×1 convolution operation, and they are input into the ReLU function as a whole together with the fifth branch to obtain features.

7. The target detection and recognition method based on adaptive search network guidance according to claim 1 is characterized in that: The multi-scale feature fusion module uses channel domain maximum pooling and channel domain average pooling operations to input the two generated compressed features into the same MLP module, then concatenates the output features, and multiplies the result obtained after Sigmoid function processing with the input features to obtain a channel attention map.

8. The target detection and recognition method based on adaptive search network guidance according to claim 7 is characterized in that: The specific operation process of the multi-scale feature fusion module is shown in the following formula: M c (F)=σ(M LP (P Avg (F));M LP (P Max (F))) in: M c (F) indicates output; F is the input feature; σ is the Sigmoid function; M LP Represents the MLP module, which is a multi-layer perceptron consisting of a convolutional layer and a ReLU activation function; P Avg With P Max They represent channel domain average pooling and channel domain maximum pooling respectively; Represents a concatenation operation.

9. The method for target detection and recognition based on adaptive search network guidance according to any one of claims 1 to 8, characterized in that: The detection network is used to detect and identify targets on images acquired by the sensor, including: Step S1: sending the image acquired by the sensor to the detection network; Step S2: The first feature extraction module captures the target feature points, and then the adaptive attention feature screening module selects the target features of different importance, deeply mines the low-level, high-level, texture, semantic and other feature information of the target's intrinsic relevance and collects them to obtain the mapping feature vector Step S3: Performed simultaneously with step S2, the search network SANet based on the dual-channel receptive field MRF module actively searches for the target area, guides the detection network to learn features, and obtains the mapping feature vector Step S4: Map feature vector and the mapping feature vector Perform point multiplication to obtain the weighted feature vector Step S5: The features outputted from step S4 are then captured by the second feature extraction module to capture the target feature points, and the adaptive attention feature screening module is used to screen out target features of different importance levels to obtain the feature vector Step S6: Performed simultaneously with step S5, the search network SANet based on the dual-channel receptive field MRF module actively searches for the target area, guides the detection network to learn features, and obtains the mapping feature vector Step S7: Map feature vector and the mapping feature vector Perform point multiplication to obtain the weighted feature vector Step S8: The features outputted from step S7 are then passed through the third feature extraction module to capture the target feature points, and the adaptive attention feature screening module is used to screen out target features of different importance levels, so as to further capture richer feature information vectors. Step S9: Performed simultaneously with step S7, the search network SANet based on the dual-channel receptive field MRF module actively searches for the target area, guides the detection network to learn features, and obtains a new mapping feature vector Step S10: Map feature vector and the mapping feature vector Perform point multiplication to obtain the weighted feature vector Step S11: The features outputted from step S4, step S7 and step S10 are respectively subjected to a channel attention operation composed of maximum pooling and average pooling, and the two generated compressed features are inputted into the same MLP module, and then the output features are concatenated, and the result obtained after Sigmoid function processing is multiplied by the input feature to obtain three channel attention maps of different scales; Step S12: further fuse the multi-scale features obtained in step S11 to obtain the final comprehensive detection and recognition result.

Images