Target detection method and device, equipment, medium and program product

Abstract

The invention provides a target detection method which can be applied to the fields of big data and artificial intelligence. The target detection method comprises the steps of performing feature extraction on a to-be-detected image through a first feature extraction network to obtain a multi-scale intermediate feature map, performing multi-scale information fusion and feature multiplexing enhancement on the multi-scale intermediate feature map to obtain a high-dimensional feature map, and performing attention enhancement processing on the high-dimensional feature map to obtain an enhanced feature map; auxiliary feature extraction is carried out on the to-be-detected image and the enhanced feature map through a second feature extraction network to obtain an auxiliary feature map, and a processing path of the second feature extraction network is determined based on the input features; and inputting the enhanced feature map into the first detection head, inputting the auxiliary feature map into the second detection head for prediction, and fusing prediction results of the first detection head and the second detection head to obtain a target detection result. The invention further provides a target detection device and equipment, a medium and a program product.

Description

Technical Field

[0001] This application relates to the fields of big data and artificial intelligence, specifically to a target detection method, apparatus, device, medium, and program product. Background Technology

[0002] Currently, object detection technology has been widely applied in various fields such as security, transportation, and healthcare, achieving excellent results. Several technology companies, research institutions, and universities have also launched various deep learning-based object detection models. These models perform exceptionally well on general datasets, exhibiting high detection accuracy and inference speed. Numerous practical applications demonstrate that deep learning object detection technology can effectively identify multiple target objects in images and supports real-time processing of large-scale video streams, demonstrating strong technological maturity and engineering feasibility.

[0003] However, existing target detection methods suffer from problems such as insufficient multi-scale target detection accuracy, weak adaptability to complex environments (such as target occlusion, lighting changes, pose differences, etc.), and poor feature representation and scene adaptability in complex scenarios such as bank risk control. As a result, the accuracy and robustness of target detection are difficult to meet the requirements. Summary of the Invention

[0004] In view of the above problems, this application provides target detection methods, apparatus, equipment, media and program products that improve the accuracy and robustness of target detection.

[0005] According to a first aspect of this application, a target detection method is provided, comprising: extracting features from an image to be detected using a first feature extraction network to obtain a multi-scale intermediate feature map; performing multi-scale information fusion and feature reuse enhancement on the multi-scale intermediate feature map to obtain a high-dimensional feature map; and performing attention enhancement processing on the high-dimensional feature map to obtain an enhanced feature map; extracting auxiliary features from the image to be detected and the enhanced feature map using a second feature extraction network to obtain an auxiliary feature map, wherein the processing path of the second feature extraction network is determined based on the input features, and the auxiliary feature map represents a multi-modal composite feature that fuses original image details and scene adaptation features; inputting the enhanced feature map into a first detection head and the auxiliary feature map into a second detection head for prediction; and fusing the prediction results of the first and second detection heads to obtain a target detection result.

[0006] According to an embodiment of this application, multi-scale information fusion and feature reuse enhancement are performed on multi-scale intermediate feature maps to obtain high-dimensional feature maps, including: compressing the channel dimension of the multi-scale intermediate feature maps to obtain compressed feature maps; expanding the receptive field of the compressed feature maps by using multiple cascaded dilated convolutional layers with sequentially increasing dilation rates to obtain the output features of each dilated convolutional layer, wherein each dilated convolutional layer receives the output features of the preceding layer as input; and generating high-dimensional feature maps based on the output features of each dilated convolutional layer.

[0007] According to an embodiment of this application, generating a high-dimensional feature map based on the output features of each dilated convolutional layer includes: concatenating and compressing the output features of each dilated convolutional layer according to the channel dimension to obtain a preliminary integrated multi-scale feature map; and fusing the preliminary integrated multi-scale feature map with the compressed feature map to obtain a high-dimensional feature map.

[0008] According to an embodiment of this application, the auxiliary feature map is obtained by using a second feature extraction network to perform auxiliary feature extraction on the image to be detected and the enhanced feature map. This includes: fusing the features of the image to be detected and the enhanced feature map to obtain an initial fused feature map; extracting fine-grained texture features and overall semantic information from the initial fused feature map, mapping the fine-grained texture features and overall semantic information to the same channel dimension to obtain a composite feature map; and using processing branches of different complexity to extract and fuse features from the composite feature map to obtain an auxiliary feature map. The weighted fusion ratio between processing branches of different complexity is determined based on the complexity of the composite feature map.

[0009] According to the embodiments of this application, feature extraction and fusion of composite feature maps using processing branches of different complexities to obtain auxiliary feature maps include: performing sparsification processing on the composite feature maps to obtain global feature maps, which represent the main patterns and key information of the composite features; performing local attention enhancement on the composite feature maps to obtain local feature maps, which represent the local details of the composite feature maps; and performing weighted fusion of the global feature maps and local feature maps with a weighted fusion ratio to obtain auxiliary feature maps.

[0010] According to an embodiment of this application, local attention enhancement is performed on a composite feature map to obtain a local feature map, which includes: obtaining spatial attention weights and channel attention weights; adaptively fusing the channel attention weights and spatial attention weights, enhancing the response differences through a nonlinear activation function, and combining the fused weights with the composite feature map to form a local feature map.

[0011] According to an embodiment of this application, obtaining spatial attention weights includes: performing feature statistics on local windows at each spatial location in the composite feature map, constructing an energy function based on the statistical results, the energy function being used to measure the feature response intensity of the local region; and normalizing and activating the energy function to obtain the spatial attention weights.

[0012] According to the embodiments of this application, a first feature extraction network, a second feature extraction network, a first detection head, and a second detection head constitute an object detection model; the training process of the object detection model includes: determining the classification loss, bounding box regression loss, and object confidence loss of the first detection head and the second detection head respectively; and weighting and fusing the classification loss, bounding box regression loss, and object confidence loss of the first detection head and the second detection head according to a preset weight ratio to form the total loss function in the training process of the object detection model.

[0013] A second aspect of this application provides a target detection apparatus, comprising: a first feature extraction module, configured to extract features from an image to be detected using a first feature extraction network to obtain a multi-scale intermediate feature map, perform multi-scale information fusion and feature reuse enhancement on the multi-scale intermediate feature map to obtain a high-dimensional feature map, and perform attention enhancement processing on the high-dimensional feature map to obtain an enhanced feature map; a second feature extraction module, configured to extract auxiliary features from the image to be detected and the enhanced feature map using a second feature extraction network to obtain an auxiliary feature map, wherein the processing path of the second feature extraction network is determined based on the input features, and the auxiliary feature map represents a multi-modal composite feature that fuses original image details and scene adaptation features; and a detection result determination module, configured to input the enhanced feature map into a first detection head and the auxiliary feature map into a second detection head for prediction, and fuse the prediction results of the first and second detection heads to obtain a target detection result.

[0014] A third aspect of this application provides an electronic device comprising: one or more processors; and a memory for storing one or more computer programs, wherein the one or more processors execute the one or more computer programs to implement the steps of the method described above.

[0015] A fourth aspect of this application also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0016] The fifth aspect of this application also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the above-described method. Attached Figure Description

[0017] The above-mentioned contents, other objects, features and advantages of this application will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0018] Figure 1 The illustrations depict application scenarios of target detection methods, apparatus, devices, media, and program products according to embodiments of this application.

[0019] Figure 2 A flowchart illustrating a target detection method according to an embodiment of this application is shown schematically.

[0020] Figure 3 The diagram illustrates the architecture of an object detection model according to an embodiment of this application.

[0021] Figure 4 The diagram schematically illustrates the architecture of a high-density path network for an object detection model according to an embodiment of this application.

[0022] Figure 5 This schematically illustrates the architecture of an adaptive timing controller for a target detection model according to an embodiment of this application;

[0023] Figure 6 The diagram illustrates the architecture of the local attention compensation mechanism of the object detection model according to an embodiment of this application.

[0024] Figure 7 This schematic diagram illustrates the structural block diagram of a target detection device according to an embodiment of the present application;

[0025] Figure 8 A block diagram schematically illustrates an electronic device suitable for implementing a target detection method according to an embodiment of this application. Detailed Implementation

[0026] The embodiments of this application will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of this application. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of this application for ease of explanation. However, it will be apparent that one or more embodiments may be implemented without these specific details. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concepts of this application.

[0027] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The terms “comprising,” “including,” etc., as used herein indicate the presence of features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0028] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0029] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0030] In the technical solution of this application, the user information (including but not limited to user personal information, user image information, user device information, such as location information) and data (including but not limited to data used for analysis, stored data, and displayed data) involved are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of related data all comply with relevant laws, regulations, and standards, take necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation entry points for users to choose to authorize or refuse.

[0031] In scenarios involving automated decision-making using personal information, the methods, devices, and systems provided in this application all offer users corresponding entry points for choosing to agree to or reject the automated decision-making results. If the user chooses to reject, the process proceeds to the expert decision-making stage. Here, "automated decision-making" refers to the activity of automatically analyzing and evaluating an individual's behavioral habits, interests, or economic, health, and credit status through computer programs, and then making a decision. Here, "expert decision-making" refers to the activity of making decisions by personnel who specialize in a particular field, possess specialized experience, knowledge, and skills, and have reached a certain level of professional expertise.

[0032] Figure 1 The illustration shows an application scenario of the target detection method, apparatus, device, medium, and program product according to embodiments of this application.

[0033] like Figure 1 As shown, application scenario 100 according to this embodiment may include a first terminal device 101 (e.g., a bank self-service terminal), a second terminal device 102 (e.g., a bank staff office terminal), a third terminal device 103 (e.g., a customer mobile terminal), a network 104, and a server 105 (e.g., a bank risk control back-end server). The network 104 serves as a medium for providing communication links between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. The network 104 may include various connection types, such as internal bank wired communication links, dedicated financial wireless communication links, or encrypted fiber optic cables, etc., to ensure the security and stability of data transmission and comply with financial industry data compliance requirements.

[0034] Users (such as bank staff and customers) can use the first terminal device 101, the second terminal device 102, and the third terminal device 103 to interact with the server 105 via the network 104 to upload images to be detected (such as ID cards and facial images) and receive target detection results (such as document authenticity determination and face-to-document matching degree). Various bank-specific client applications can be installed on the first terminal device 101, the second terminal device 102, and the third terminal device 103, such as "document verification application" for bank self-service terminals, "risk control and audit application" for staff office terminals, and "remote account opening application" for customer mobile terminals (for example only). These applications integrate image acquisition modules (such as cameras) to acquire images of the targets to be detected in real time.

[0035] The first terminal device 101, the second terminal device 102, and the third terminal device 103 can be various electronic devices adapted to bank risk control scenarios: the first terminal device 101 (bank self-service terminal) can be a vertical terminal with a high-definition camera and a touch screen, supporting customers to scan their ID cards and collect their faces independently; the second terminal device 102 (bank staff office terminal) can be a desktop computer or portable laptop with an external high-speed scanner, used by staff to manually verify customer documents; the third terminal device 103 (customer mobile terminal) can be a smartphone or tablet computer that supports high-definition front / rear cameras, meeting the needs of customers to remotely submit documents and facial information.

[0036] Server 105 can be a bank-dedicated risk control backend server used to deploy the target detection algorithm of this application embodiment, providing target detection services for images uploaded by the first terminal device 101, the second terminal device 102, and the third terminal device 103. Server 105 can perform multi-scale feature extraction, attention enhancement, auxiliary feature fusion, and collaborative prediction on the received image data to be detected, and feed back the processing results to the corresponding terminal devices.

[0037] It should be noted that the target detection algorithm provided in this application embodiment can generally be executed by server 105. Accordingly, the core modules of the algorithm can generally be located in server 105, while the terminal device is responsible for image acquisition and result display, reducing terminal computing power consumption. In some low-latency scenarios, the target detection algorithm provided in this application embodiment can also be executed locally by the terminal device, or executed collaboratively by server 105 and the terminal device, to balance real-time performance and computing power costs. Accordingly, some lightweight modules of the algorithm can also be located in the terminal device, while the core complex modules are still deployed on server 105.

[0038] It should be understood that Figure 1The number of terminal devices, networks, and servers shown is merely illustrative. Depending on the actual scale of a bank's business, it can have any number of terminal devices, multiple layers of networks, and server clusters. For example, large banks can deploy multiple regional server nodes to process detection tasks uploaded by terminal devices locally, further reducing data transmission latency and improving risk control efficiency.

[0039] It should be noted that the target detection method and apparatus of this application are deeply adaptable to the risk control needs of banks in the fintech field, and are particularly suitable for business systems such as remote account opening, self-service terminal document verification, and branch customer behavior monitoring. In bank risk control scenarios, documents and faces are the core objects of customer identity verification and risk identification. Their accurate identification and real-time verification are directly related to the authenticity of customer identity, business compliance, and fund security. They need to cope with the target detection requirements in complex environments (lighting changes, angle tilt, occlusion and reflection) and strict privacy protection requirements (such as local processing of sensitive information). This method effectively solves several key technical bottlenecks faced by existing target detection models in bank risk control image recognition tasks through the core design of high-density path network multi-scale fusion, adaptive dynamic routing, and local attention enhancement.

[0040] It should be noted that the target detection method and apparatus of this application can be used in bank risk control in the financial technology field, and can also be used in any field other than the financial field, such as abnormal behavior detection of personnel in the security field, merchandise display recognition in the retail field, and medical record and document classification in the medical field. The application field of the target detection method and apparatus of this application is not limited.

[0041] The following will be based on Figure 1 The described scene, through Figures 2-6 The target detection method according to the embodiments of this application will be described in detail.

[0042] Figure 2 A flowchart illustrating a target detection method according to an embodiment of this application is shown.

[0043] like Figure 2 As shown, the target detection method of this embodiment includes operations S210 to S230, and the target detection method can be executed by server 105.

[0044] In operation S210, the first feature extraction network extracts features from the image to be detected to obtain a multi-scale intermediate feature map. Multi-scale information fusion and feature reuse enhancement are performed on the multi-scale intermediate feature map to obtain a high-dimensional feature map. Attention enhancement processing is then performed on the high-dimensional feature map to obtain an enhanced feature map.

[0045] In operation S220, the second feature extraction network performs auxiliary feature extraction on the image to be detected and the enhanced feature map to obtain an auxiliary feature map. The processing path of the second feature extraction network is determined based on the input features. The auxiliary feature map represents a multimodal composite feature that integrates the original image details and scene adaptation features.

[0046] In operation S230, the enhanced feature map is input into the first detection head and the auxiliary feature map is input into the second detection head for prediction. The prediction results of the first and second detection heads are fused to obtain the target detection result.

[0047] As an example, server 105 can receive the image to be detected provided by the user (such as an image of an ID card and a face in the same frame), and first perform standardization and size adaptation, for example, normalizing the pixel values ​​to [0,1] to maintain the original resolution and avoid stretching, and then improving the unevenness of illumination through adaptive histogram equalization; the image quality can be evaluated by the Laplacian variance method, and if the blur index or the lack of effective edge features, it will return prompts such as "image quality is unqualified" or "no detection target" respectively. Sensitive information such as faces and ID numbers in the image to be detected can be desensitized locally during the preprocessing stage.

[0048] Multi-scale intermediate feature maps can represent the shallow and mid-level basic features of the image to be detected. High-dimensional feature maps can simultaneously contain small-scale details and large-scale semantics, and enhanced feature maps can be more focused on the core detection object through attention enhancement. The first feature extraction network can be the feature recognition backbone network in the object detection model. The first feature extraction network can sequentially perform feature extraction, multi-scale information fusion, and feature reuse enhancement operations on the image to be detected to obtain an enhanced feature map. The second feature extraction network can be a branch auxiliary detection network running in parallel with the first feature extraction network. The first detection head can be the main detection head corresponding to the feature recognition backbone network, and the second detection head can be the auxiliary detection head corresponding to the branch auxiliary detection network. The feature recognition backbone network, branch auxiliary detection network, main detection head, and auxiliary detection head can constitute the object detection model. That is to say, the object detection methods operating S210 to S230 can be implemented through the object detection model. Figure 3 The diagram illustrates the architecture of an object detection model according to an embodiment of this application.

[0049] like Figure 3As shown, in the target detection model of this embodiment, the feature recognition backbone network can use multiple alternating convolutional layers, neural network blocks, and downsampling to extract features from the image to be detected, obtaining multi-scale intermediate feature maps. Then, the multi-scale intermediate feature maps are input into the HyperDensePathNet (HDPN) module designed in this application for multi-scale information fusion and feature reuse enhancement, resulting in high-dimensional feature maps. The HDPN module can achieve efficient and powerful context awareness by combining multi-scale dilated convolution with a forward dense connection mechanism and introducing a lightweight channel compression bypass branch (Cross Stage Partial, CSP). Next, the high-dimensional feature maps are input into the Focus Path Attention (FPA) mechanism designed in this application for attention enhancement processing, resulting in enhanced feature maps P3, P4, and P5. Feature map P3 has high resolution and a small number of channels, representing fine-grained details at the lower level, such as text on the document and local facial features. P4 has medium resolution and a medium number of channels, representing mid-level target features, such as the complete document outline and the overall shape of the face. P5 has low resolution and a high number of channels, representing high-level global semantic information, such as the spatial relationship between the face and the document and the background environment. The branch-assisted detection network can simultaneously receive the image to be detected and the enhanced feature maps P3, P4, and P5. Auxiliary feature extraction is performed using the Adaptive SpatioTemporal Modulator (ASTM) module designed in this application, resulting in auxiliary feature maps P3', P4', and P5'. The ASTM processing path can be dynamically optimized based on the input features, enabling the auxiliary feature maps to fuse original image details with scene-adaptive features. The ASTM module can improve the target detection model's ability to model long-distance dependencies and achieve joint modeling of global semantic information and local detail features. Next, the enhanced feature maps P3, P4, P5 and the auxiliary feature maps P3', P4', P5' are input into their respective main detection head and auxiliary detection head for regression type prediction, and the prediction results are fused to obtain the final target detection result. The auxiliary detection head can be located in the middle layer of the network, aiming to guide the model to learn richer feature representations and provide richer gradient information to assist the main detection head in optimization. Specifically, the auxiliary detection head focuses more on improving sample recall, while the main detection head is responsible for filtering high-quality prediction results. In some embodiments, the specific structures of the HDPN, ASTM, and FPA modules can be as follows: Figure 3 , Figure 4 , Figure 5 As shown in the figures, these modules will be described in detail below with reference to the accompanying figures.

[0050] In existing target detection methods for banking and financial risk control scenarios, document-type targets are often small in size and have blurred edges, resulting in poor recognition performance. For example, during the process of a customer submitting their ID card, the image may be tilted, blurred, or even partially occluded due to factors such as shooting angle and lighting conditions. Secondly, factors such as changes in lighting, occlusion, and reflection affect the stability of the model, causing large fluctuations in recognition results and making it easy to have false positives or false negatives. Thirdly, there is a lack of effective multimodal fusion mechanisms, and relying solely on static images makes it difficult to capture the customer's true emotions and behavioral states. For example, certain abnormal behaviors are accompanied by specific micro-expressions, hesitant movements, and other abnormal signals, but most current systems have not yet incorporated such unstructured information into their analysis. In addition, most existing methods have not been optimized for lightweight banking terminal devices, and problems such as insufficient computing power and response latency often occur during deployment, limiting their application in actual business. Especially in resource-constrained environments such as mobile clients and self-service terminals, traditional models often fail to meet the requirements of low power consumption, low latency, and high concurrency. Meanwhile, some systems lack localized deployment and data anonymization mechanisms when processing sensitive information such as faces and identification documents, which does not meet the privacy protection and compliance requirements of the financial industry.

[0051] The target detection method and model in this embodiment integrate features at different resolutions, preserving details of small targets such as document text and facial micro-expressions, while also associating spatial semantics such as faces and documents. This solves the problem of lost details or semantic fragmentation in traditional single-scale features, laying the foundation for improved accuracy. Attention focus further filters key regions, weakens background interference, reduces false detections in complex environments, and enhances accuracy. Dynamic assistance adapts the processing path according to the input complexity. For example, low-complexity samples use a lightweight branch to adapt to the low computing power of the terminal, while high-complexity samples use a high-precision branch to ensure the detection quality on the server side, achieving a balance between efficiency and hardware compatibility. Dual detection head collaboration allows the main detection head to ensure high-quality predictions, while the auxiliary detection head supplements missed targets, improving the reliability of the results. The entire chain requires no additional data or complex post-processing, remains stable under scenarios such as changes in lighting and occlusion, accurately adapts to the needs of bank risk control, and can be widely used in various tasks such as document recognition, liveness detection, and behavior analysis.

[0052] In some embodiments, the HDPN module can perform channel compression, multi-scale dilated convolution, and feature fusion. Specifically, the channel dimension of the multi-scale intermediate feature maps can be compressed first to obtain compressed feature maps. For example, the number of channels in feature maps P3, P4, and P5 can be compressed to half of their original dimensions using 1×1 convolutions, reducing subsequent computational complexity. Then, multiple cascaded dilated convolutional layers with progressively increasing dilation rates are used to expand the receptive field of the compressed feature maps, with each dilated convolutional layer receiving the output features of the preceding layer as input, forming a cascaded feature reuse mechanism. For example, four cascaded dilated convolutional layers can be used. Finally, a high-dimensional feature map is generated based on the output features of each dilated convolutional layer.

[0053] This embodiment enhances the multi-scale expressive power of features while reducing computational load through channel compression, multi-scale dilated convolution, and cascaded feature reuse, thereby improving the feature capture effect for targets of different sizes.

[0054] In some embodiments, the HDPN module can concatenate and compress the output features of each dilated convolutional layer along the channel dimension to obtain preliminary integrated multi-scale features. Then, the preliminary integrated multi-scale features are fused with the compressed feature map (lightweight bypass branch) to obtain a high-dimensional feature map. The high-dimensional feature map can simultaneously contain small-scale details and large-scale semantics. This embodiment, by fusing the lightweight bypass branch with the main path features, preserves the complementary information of low-level details and high-level semantics, enhancing the integrity and robustness of the high-dimensional feature map.

[0055] Figure 4 The schematic diagram illustrates the architecture of a high-density path network (HDPN) for object detection according to an embodiment of this application.

[0056] like Figure 4 As shown, in this embodiment, the input layer of the HDPN module first performs channel compression on the multi-scale intermediate features of the input through 1×1 convolution. For example, when the number of input channels is 512, it is compressed to 256 to reduce the computational complexity of subsequent steps. Then, four cascaded dilated convolutional layers are used, with dilation rates of 3, 6, 12, and 18 respectively, a kernel size of 3×3, and a stride of 1, progressively expanding the receptive field. These layers receive the outputs of all preceding layers as input through dense forward connections, forming a cascaded feature reuse mechanism. Next, the outputs of each layer are concatenated and then unified by a 1×1 convolution to achieve the same channel dimension. This concatenation is then performed with the lightweight CSP bypass branch (i.e., the compressed multi-scale intermediate feature map) channel-wise (concatenate), and finally unified to the target number of channels through a 1×1 convolution, outputting a high-dimensional feature map with strong expressive power. This design significantly enhances the model's adaptability to complex environmental changes and improves the overall robustness of detection.

[0057] To further enhance the target detection model's ability to model long-range dependencies and to achieve joint modeling of global semantic information and local detail features of images, this application introduces an ASTM structure after the network's neck module.

[0058] In some embodiments, the ASTM module can fuse the image to be detected and the enhanced feature map to obtain an initial fused feature map. Then, fine-grained texture features and overall semantic information are extracted from the initial fused feature map. After mapping the fine-grained texture features and overall semantic information to the same channel dimension, a composite feature map is formed. Next, different complexity processing branches are used to extract and fuse features from the composite feature map to obtain an auxiliary feature map. The weighted fusion ratio between different complexity processing branches can be determined by evaluating the complexity of the composite feature map. Understandably, branches with high complexity processing capabilities can be used to process composite feature maps with higher complexity, and branches with low complexity processing capabilities can be used to process composite feature maps with lower complexity, in order to improve processing efficiency and save computational resources.

[0059] Figure 5 The diagram illustrates the architecture of an adaptive time-limited controller for a target detection model according to an embodiment of this application.

[0060] like Figure 5 As shown, in this embodiment, the ASTM module can include three key components: a hierarchical state-space modeling unit, a dynamic routing controller, and a local attention compensation mechanism. The hierarchical state-space modeling unit captures fine-grained texture features and overall semantic information through parallel local state-space model (SSM) branches and a global SSM branch, respectively, and fuses the outputs of both using a learnable linear projection matrix to form a composite feature representation. The local SSM branch uses a smaller time step (e.g., Δ=0.1) to focus on modeling details such as edges and micro-expressions; the global SSM branch uses a larger time step (e.g., Δ=1.0) to extract the overall structure and long-range dependencies of the image. This dual-path structure allows the model to maintain high-resolution detail perception while accurately understanding the overall semantics of the image, thereby improving the accuracy of target recognition. This embodiment balances the accuracy and efficiency of feature extraction through feature alignment fusion and dynamic complexity adaptation branching, improving the model's adaptability to complex scenes.

[0061] In some embodiments, the ASTM module can sparsify the composite feature map to obtain a global feature map, which represents the main patterns and key information of the composite feature. Then, local attention enhancement is applied to the composite feature map to obtain a local feature map, which represents the local details of the composite feature map. Finally, the global and local feature maps are weighted and fused using a weighted fusion ratio to obtain an auxiliary feature map.

[0062] like Figure 5 As shown, considering the stringent requirements for inference efficiency and power consumption control during actual deployment, this application introduces a dynamic routing controller in the ASTM module. This controller generates a gating signal based on the complexity of the input samples, dynamically controlling the weighted fusion ratio between "low-complexity experts" and "high-complexity experts." Based on the energy distribution and spatial complexity evaluation results of the input features, the controller intelligently selects the optimal computation path, significantly improving inference efficiency and reducing average power consumption without affecting detection accuracy. This mechanism is particularly suitable for deployment in edge devices or battery-powered scenarios, effectively extending device battery life and enhancing system usability. This embodiment, through weighted fusion of sparse global features and enhanced local attention, reduces redundant computation while preserving key details, optimizing the discriminative power of multi-scale auxiliary features.

[0063] To further enhance the model's ability to identify key information such as small targets and boundary regions, while avoiding the introduction of too many parameters that would increase model complexity, this application also introduces a local attention compensation mechanism (FPA) in the ASTM module.

[0064] The FPA module can perform local attention enhancement on composite feature maps to obtain local feature maps. In some embodiments, the FPA module can obtain spatial attention weights and channel attention weights. Then, the channel attention weights and spatial attention weights are adaptively fused, and the response differences are enhanced through a non-linear activation function. The fused weights are then combined with the composite feature map to form a local feature map. Through the adaptive fusion of channel and spatial attention, the response of key local features is strengthened, improving the accuracy of detail capture for small targets and occluded targets.

[0065] The FPA module can obtain spatial attention weights. In some embodiments, the FPA module can perform feature statistics on local windows at each spatial location in the composite feature map, construct an energy function based on the statistical results, and use the energy function to measure the feature response intensity of the local region. Then, the energy function is normalized and activated to obtain the spatial attention weights.

[0066] Figure 6 The diagram illustrates the architecture of the local attention compensation mechanism of the object detection model according to an embodiment of this application.

[0067] like Figure 6 As shown, the FPA mechanism can construct an energy function to measure the intensity of the feature response by statistically analyzing the mean and standard deviation of a 3×3 local window at each location. The energy function can be expressed as:

[0068]

[0069] Where E(i) is the energy function value (at the i-th position), used to describe the importance or significance of the feature at that position; V i σ is the feature value at the i-th position, which can also be used as the activation value of a channel at that spatial position; μ is the mean (local or global) of the feature map, used to describe the baseline response intensity of the current region; 2 ε is the variance of the feature map, used to measure the dispersion of the feature distribution within the region; ε is a minimal constant used to prevent division by zero errors.

[0070] Subsequently, channel attention weights (ChannelAttn) and spatial attention weights (SpatialAttn) are generated to capture significant differences between channels and in spatial location. Based on this, the two attention methods can be adaptively fused using a small number of learnable parameters α and β. The fused weights can be expressed as:

[0071]

[0072] Next, a non-linear activation function is used to enhance the response differences, thus forming the final attention map. The attention map can be represented as:

[0073] Y = Tanh[Attn(X)]

[0074] Tanh is the hyperbolic tangent function, whose core function is to perform non-linear transformation on the input features, enhance the model's ability to express complex features, and map the output values ​​to a fixed range, thereby improving training stability.

[0075] The FPA mechanism in this embodiment achieves precise focusing on key regions with almost no additional parameters, effectively improving the model's detection accuracy. It generates spatial attention weights based on local window energy distribution, accurately locating key target regions, effectively suppressing background interference, and enhancing feature focusing capabilities.

[0076] It is worth mentioning that both the feature recognition backbone network and ASTM can use the FPA mechanism to enhance the attention of the input feature map, such as... Figure 3 and Figure 5 As shown. In other words, the feature recognition backbone network and the attention mechanism in ASTM can adopt the same structure.

[0077] Please continue reading. Figure 3 The training process of the object detection model may include: determining the classification loss, bounding box regression loss, and object confidence loss of the first and second detection heads respectively; and weighting and fusing the classification loss, bounding box regression loss, and object confidence loss of the first and second detection heads according to a preset weight ratio to form the total loss function in the training process of the object detection model.

[0078] Regarding training stability and generalization ability, this application also introduces an auxiliary head mechanism in the detection head design. This auxiliary head, located in the middle layer of the network, aims to guide the model to learn richer feature representations and provide richer gradient information to assist the main detection head in optimization. Specifically, the auxiliary head focuses on improving sample recall, while the main detection head is responsible for filtering high-quality prediction results. When calculating the loss function, the losses of the auxiliary and main detection heads can be weighted and fused at a ratio of, for example, 0.25:1, thereby achieving a more stable training process and stronger model generalization ability. This design helps alleviate the gradient vanishing problem in deep model training, improving model convergence speed and final performance.

[0079] In addition, to address the common issues of class imbalance and sensitivity to localization errors in target detection tasks, this application adopts a combination of three types of loss functions in the design of the detection head loss function: classification loss, bounding box regression loss, and target confidence loss.

[0080] As an example, the classification loss can be achieved using the Binary Cross-Entropy Loss (BCELoss) function to determine whether the predicted class is correct.

[0081] As an example, the bounding box regression loss can be a combination of Distribution Focal Loss (DFL Loss) and Wise Intersection over Union Loss (WIOU Loss), where DFL Loss can alleviate the class imbalance problem. DFL Loss can be expressed as:

[0082]

[0083] Where Si is the global minimum solution, and y is the target label. i y i+1 These are the values ​​of the two anchor points adjacent to the target label y.

[0084] WIOU Loss can more accurately measure the positional deviation of prediction boxes based on a dynamic non-monotonic focusing mechanism. It adds two factors to the ordinary IOU structure: the former amplifies the prediction error of average quality predictions, and the latter reduces the penalty for high-quality predictions, thereby mitigating the differences caused by inconsistent data quality. WIOU Loss can be expressed as:

[0085] L WIOU =R WIOU × L IOU

[0086] WIOU Loss adds to the regular IOU structure and The former can amplify ordinary mass L IOU The latter can reduce the quality of R WIOU This reduces the discrepancies caused by inconsistent data quality.

[0087] As an example, the target confidence loss can be constructed based on binary cross-entropy to assess whether the predicted target exists. The aforementioned loss functions can be weighted and fused to form a total loss function, guiding the entire network training process and ensuring the model maintains good detection performance across various scenarios.

[0088] In this embodiment, the model training process uses a combination of multiple loss functions and weighted loss from dual detectors to alleviate class imbalance and localization error, thereby enhancing the model's training stability and generalization ability.

[0089] In summary, the HDPN, ASTM, FPA, and dual-detection-head structure proposed in this application constitute a highly efficient, flexible, and adaptable deep learning-based object detection framework with significant engineering application value and promising prospects for widespread adoption. Taking the banking business scenario as an example, this application demonstrates strong algorithm accuracy and generalization ability, achieving efficient identification and precise localization of target objects in various banking business scenarios. This object detection method can adapt to image input under different lighting, angle, and occlusion conditions, exhibiting good environmental robustness and wide applicability to various tasks such as document recognition, liveness detection, and behavior analysis. In terms of real-time performance, this object detection method can operate stably under existing bank video surveillance equipment conditions, possessing the ability to continuously process video streams with a smooth overall response, meeting the basic requirements for real-time object detection in daily operations and providing reliable support for the intelligent upgrade of bank security and control systems. Addressing the specific characteristics of banking business scenarios, this application has made targeted optimizations in data management, model deployment, and functional design, enhancing the system's flexibility and customizability. The system can be quickly adapted to the actual needs of different branches, possessing good promotion potential and feasibility for implementation.

[0090] Based on the above-described target detection method, this application also provides a target detection device. The following will be combined with... Figure 7 The device is described in detail.

[0091] Figure 7 A schematic block diagram of a target detection device according to an embodiment of this application is shown.

[0092] like Figure 7 As shown, the target detection device 700 of this embodiment includes a first feature extraction module 710, a second feature extraction module 720, and a detection result determination module 730.

[0093] The first feature extraction module 710 is used to extract features from the image to be detected through a first feature extraction network to obtain a multi-scale intermediate feature map, perform multi-scale information fusion and feature reuse enhancement on the multi-scale intermediate feature map to obtain a high-dimensional feature map, and perform attention enhancement processing on the high-dimensional feature map to obtain an enhanced feature map. In one embodiment, the first feature extraction module 710 can be used to perform the operation S210 described above, which will not be repeated here.

[0094] The second feature extraction module 720 is used to perform auxiliary feature extraction on the image to be detected and the enhanced feature map through the second feature extraction network to obtain an auxiliary feature map. The processing path of the second feature extraction network is determined based on the input features. The auxiliary feature map represents a multimodal composite feature that integrates the original image details and scene adaptation features. In one embodiment, the second feature extraction module 720 can be used to perform the operation S220 described above, which will not be repeated here.

[0095] The detection result determination module 730 is used to input the enhanced feature map into the first detection head and the auxiliary feature map into the second detection head for prediction, and to fuse the prediction results of the first and second detection heads to obtain the target detection result. In one embodiment, the detection result determination module 730 can be used to perform the operation S230 described above, which will not be repeated here.

[0096] According to an embodiment of this application, the first feature extraction module 710 is further used to compress the channel dimension of the multi-scale intermediate feature map to obtain a compressed feature map; to expand the receptive field of the compressed feature map by using multiple cascaded dilated convolutional layers with sequentially increasing dilation rates to obtain the output features of each dilated convolutional layer, wherein each dilated convolutional layer receives the output features of the preceding layer as input; and to generate a high-dimensional feature map based on the output features of each dilated convolutional layer.

[0097] According to the embodiments of this application, the first feature extraction module 710 is further used to splice and compress the output features of each dilated convolutional layer according to the channel dimension to obtain a preliminary integrated multi-scale feature; and to fuse the preliminary integrated multi-scale feature with the compressed feature map to obtain a high-dimensional feature map.

[0098] According to an embodiment of this application, the second feature extraction module 720 is used to fuse the image to be detected and the enhanced feature map to obtain an initial fused feature map; extract fine-grained texture features and overall semantic information from the initial fused feature map, and map the fine-grained texture features and overall semantic information to the same channel dimension to obtain a composite feature map; use processing branches of different complexity to extract and fuse features from the composite feature map to obtain an auxiliary feature map, and the weighted fusion ratio between processing branches of different complexity is determined based on the complexity of the composite feature map.

[0099] According to an embodiment of this application, the second feature extraction module 720 is used to perform sparsification processing on the composite feature map to obtain a global feature map, which represents the main pattern and key information of the composite feature map; to perform local attention enhancement on the composite feature map to obtain a local feature map, which represents the local details of the composite feature map; and to perform weighted fusion of the global feature map and the local feature map with a weighted fusion ratio to obtain an auxiliary feature map.

[0100] According to an embodiment of this application, the second feature extraction module 720 is used to obtain spatial attention weights and channel attention weights; adaptively fuse the channel attention weights and spatial attention weights, and enhance the response differences through a nonlinear activation function, and combine the fused weights with the composite feature map to form a local feature map.

[0101] According to an embodiment of this application, the second feature extraction module 720 is used to perform feature statistics on the local window at each spatial location in the composite feature map, construct an energy function based on the statistical results, and use the energy function to measure the feature response intensity of the local region; the energy function is normalized and activated to obtain the spatial attention weight.

[0102] According to the embodiments of this application, a first feature extraction network, a second feature extraction network, a first detection head, and a second detection head constitute an object detection model; the training process of the object detection model includes: determining the classification loss, bounding box regression loss, and object confidence loss of the first detection head and the second detection head respectively; and weighting and fusing the classification loss, bounding box regression loss, and object confidence loss of the first detection head and the second detection head according to a preset weight ratio to form the total loss function in the training process of the object detection model.

[0103] According to embodiments of this application, any plurality of the above modules can be combined into one module, or any one of the modules can be split into multiple modules. Alternatively, at least a portion of the functionality of one or more of these modules can be combined with at least a portion of the functionality of other modules and implemented in one module. According to embodiments of this application, at least one of the above modules can be at least partially implemented as hardware circuitry, such as a Field Programmable Gate Array (FPGA), a Programmable Logic Array (PLA), a System-on-Chip, a System-on-Substrate, a System-on-Package, an Application-Specific Integrated Circuit (ASIC), or implemented in hardware or firmware by any other reasonable means of integrating or packaging the circuitry, or implemented in any one of software, hardware, and firmware methods, or in a suitable combination of any of these. Alternatively, at least one of the above modules can be at least partially implemented as a computer program module, which, when run, can perform corresponding functions.

[0104] Figure 8 A block diagram schematically illustrates an electronic device suitable for implementing a target detection method according to an embodiment of this application.

[0105] like Figure 8 As shown, an electronic device 800 according to an embodiment of this application includes a processor 801, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 802 or a program loaded from a storage portion 808 into a random access memory (RAM) 803. The processor 801 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 801 may also include onboard memory for caching purposes. The processor 801 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of this application.

[0106] RAM 803 stores various programs and data required for the operation of electronic device 800. Processor 801, ROM 802, and RAM 803 are interconnected via bus 804. Processor 801 executes various operations of the method flow according to embodiments of this application by executing programs in ROM 802 and / or RAM 803. It should be noted that programs may also be stored in one or more memories other than ROM 802 and RAM 803. Processor 801 may also execute various operations of the method flow according to embodiments of this application by executing programs stored in one or more memories.

[0107] According to embodiments of this application, the electronic device 800 may further include an input / output (I / O) interface 805, which is also connected to a bus 804. The electronic device 800 may also include one or more of the following components connected to the input / output (I / O) interface 805: an input section 806 including a keyboard, mouse, etc.; an output section 807 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 808 including a hard disk, etc.; and a communication section 809 including a network interface card such as a LAN card, modem, etc. The communication section 809 performs communication processing via a network such as the Internet. A drive 810 is also connected to the input / output (I / O) interface 805 as needed. A removable medium 811, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 810 as needed so that computer programs read from it can be installed into the storage section 808 as needed.

[0108] This application also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of this application.

[0109] According to embodiments of this application, the computer-readable storage medium can be a non-volatile computer-readable storage medium, such as including but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to embodiments of this application, the computer-readable storage medium may include ROM 802 and / or RAM 803 and / or one or more memories other than ROM 802 and RAM 803 described above.

[0110] Embodiments of this application also include a computer program product comprising a computer program containing program code for performing the methods shown in the flowchart. When the computer program product is run on a computer system, the program code enables the computer system to implement the target detection method provided in the embodiments of this application.

[0111] When the computer program is executed by the processor 801, it performs the functions defined in the system / apparatus of this application embodiment. According to the embodiments of this application, the systems, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0112] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and may be downloaded and installed via the communication section 809, and / or installed from a removable medium 811. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0113] In such an embodiment, the computer program can be downloaded and installed from a network via the communication section 809, and / or installed from the removable medium 811. When the computer program is executed by the processor 801, it performs the functions defined in the system of this application embodiment. According to the embodiments of this application, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0114] According to embodiments of this application, program code for executing the computer programs provided in the embodiments of this application can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, languages ​​such as Java, C++, Python, "C", or similar programming languages. The program code can be executed entirely on the user's computing device, partially on the user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0115] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0116] Those skilled in the art will understand that the features described in the various embodiments of this application can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, the features described in the various embodiments of this application can be combined and / or combined in various ways without departing from the spirit and teachings of this application. All such combinations and / or combinations fall within the scope of this application.

Claims

1. A target detection method, characterized in that, The method includes: The first feature extraction network extracts features from the image to be detected to obtain a multi-scale intermediate feature map. The multi-scale intermediate feature map is then subjected to multi-scale information fusion and feature reuse enhancement to obtain a high-dimensional feature map. Finally, the high-dimensional feature map is subjected to attention enhancement processing to obtain an enhanced feature map. The second feature extraction network performs auxiliary feature extraction on the image to be detected and the enhanced feature map to obtain an auxiliary feature map. The processing path of the second feature extraction network is determined based on the input features. The auxiliary feature map represents a multimodal composite feature that integrates the original image details and scene adaptation features. The enhanced feature map is input into the first detection head, and the auxiliary feature map is input into the second detection head for prediction. The prediction results of the first detection head and the second detection head are then fused to obtain the target detection result.

2. The target detection method according to claim 1, characterized in that, The step of performing multi-scale information fusion and feature reuse enhancement on the multi-scale intermediate feature map to obtain a high-dimensional feature map includes: The multi-scale intermediate feature map is compressed in terms of channel dimension to obtain a compressed feature map. The receptive field of the compressed feature map is expanded by using multiple cascaded dilated convolutional layers with sequentially increasing dilation rates to obtain the output features of each dilated convolutional layer. Each dilated convolutional layer receives the output features of the preceding layer as input. The high-dimensional feature map is generated based on the output features of each of the dilated convolutional layers.

3. The target detection method according to claim 2, characterized in that, The generation of the high-dimensional feature map based on the output features of each of the dilated convolutional layers includes: The output features of each of the dilated convolutional layers are spliced ​​and compressed according to the channel dimension to obtain preliminary integrated multi-scale features; The initially integrated multi-scale features are fused with the compressed feature map to obtain the high-dimensional feature map.

4. The target detection method according to claim 1, characterized in that, The step of performing auxiliary feature extraction on the image to be detected and the enhanced feature map through the second feature extraction network to obtain the auxiliary feature map includes: The image to be detected and the enhanced feature map are fused to obtain an initial fused feature map; Fine-grained texture features and overall semantic information are extracted from the initial fused feature map, and the fine-grained texture features and overall semantic information are mapped to the same channel dimension to obtain a composite feature map; The composite feature map is processed and fused using processing branches of different complexities to obtain the auxiliary feature map. The weighted fusion ratio between the processing branches of different complexities is determined based on the complexity of the composite feature map.

5. The target detection method according to claim 4, characterized in that, The auxiliary feature map is obtained by extracting and fusing features from the composite feature map using processing branches of different complexities, including: The composite feature map is sparsified to obtain a global feature map, which represents the main patterns and key information of the composite features. The composite feature map is subjected to local attention enhancement to obtain a local feature map, which represents the local details of the composite feature map; The global feature map and the local feature map are weighted and fused according to the weighted fusion ratio to obtain the auxiliary feature map.

6. The target detection method according to claim 5, characterized in that, The process of performing local attention enhancement on the composite feature map to obtain a local feature map includes: Obtain spatial attention weights and channel attention weights; The channel attention weights and spatial attention weights are adaptively fused, and the response differences are enhanced by a nonlinear activation function. The fused weights are then combined with the composite feature map to form the local feature map.

7. The target detection method according to claim 6, characterized in that, The obtained spatial attention weights include: Feature statistics are performed on the local windows at each spatial location in the composite feature map, and an energy function is constructed based on the statistical results. The energy function is used to measure the feature response intensity of the local region. The energy function is normalized and activated to obtain the spatial attention weights.

8. The target detection method according to claim 1, characterized in that, The first feature extraction network, the second feature extraction network, the first detection head, and the second detection head constitute a target detection model; the training process of the target detection model includes: Determine the classification loss, bounding box regression loss, and target confidence loss for the first detection head and the second detection head, respectively. According to a preset weight ratio, the classification loss, bounding box regression loss, and target confidence loss of the first detection head and the second detection head are weighted and fused to form the total loss function in the training process of the target detection model.

9. A target detection device, characterized in that, The device includes: The first feature extraction module is used to extract features from the image to be detected through the first feature extraction network to obtain a multi-scale intermediate feature map, perform multi-scale information fusion and feature reuse enhancement on the multi-scale intermediate feature map to obtain a high-dimensional feature map, and perform attention enhancement processing on the high-dimensional feature map to obtain an enhanced feature map. The second feature extraction module is used to perform auxiliary feature extraction on the image to be detected and the enhanced feature map through the second feature extraction network to obtain an auxiliary feature map. The processing path of the second feature extraction network is determined based on the input features. The auxiliary feature map represents a multimodal composite feature that integrates the original image details and scene adaptation features. The detection result determination module is used to input the enhanced feature map into the first detection head and the auxiliary feature map into the second detection head for prediction, and to fuse the prediction results of the first detection head and the second detection head to obtain the target detection result.

10. An electronic device, comprising: One or more processors; Memory, used to store one or more computer programs. The characteristic feature is that the one or more processors execute the one or more computer programs to implement the steps of the method according to any one of claims 1 to 8.

11. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 8.

12. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 8.

Images