Target detection methods, devices, equipment and storage media in wide-area complex scenarios

Abstract

This application relates to a method, apparatus, device, and storage medium for target detection in wide-area complex scenes. The method includes: extracting multi-level features from a wide-area complex scene image using a backbone network to obtain low-level features, mid-level features, and high-level features; encoding the high-level features using an EDFPT module to obtain encoded features; inputting the low-level features, mid-level features, and encoded features into a feature fusion module to obtain fused features; using an IoU-aware query selection strategy to select a fixed number of image features from the fused features to obtain an initial query vector; and processing the initial query vector using a decoder with an auxiliary prediction head to obtain the target detection result for the wide-area complex scene image. This method has higher feature sensitivity for dense small targets and irregularly shaped targets in wide-area complex scene images, effectively improving detection accuracy while maintaining a lightweight model, and reducing the false negative rate compared to RT-DETR.

Description

Technical Field

[0001] This invention belongs to the field of target detection technology, and relates to a target detection method, device, equipment and storage medium in a wide range of complex scenes. Background Technology

[0002] 3D scene rendering for wide-area complex scenes has become a hot topic in interdisciplinary research, integrating knowledge from computer vision, computer graphics, and photogrammetry. In the visual domain of wide-area complex scenes, the integrated application of sensors such as LiDAR and high-resolution cameras shows great promise. With the development of computer technology, wide-area complex scene rendering technology is of vital importance in fields such as geographic mapping and traffic monitoring. Object detection in wide-area complex scene images is an important branch of computer vision, aiming to automatically detect targets of interest from wide-area complex scene images. However, wide-area complex scene images are characterized by complex backgrounds and dense occlusion, objects are typically small in size, and objects may occlude each other, increasing the difficulty of feature extraction. Furthermore, the random orientation and diverse shapes of targets, along with the significant multi-scale characteristics of objects in complex scenes, further increase the difficulty of detection. To address these challenges, researchers are increasingly leveraging traditional image processing techniques, machine learning, and deep learning to improve the accuracy and efficiency of object detection in wide-area complex scenes.

[0003] In recent years, data-driven methods such as Neural Radiation Field (NeRF) and 3D Gaussian Scattering (3DGS) have emerged, revolutionizing the field of 3D reconstruction of wide-area complex scenes. NeRF achieves high-quality synthesis of new perspectives by modeling the scene as a continuous radiation field, but its training is time-consuming. In contrast, 3DGS uses a 3D Gaussian distribution to simultaneously model geometric structure and appearance features, achieving not only efficient and high-fidelity reconstruction and rendering but also significant computational advantages over NeRF. In wide-area complex scene object detection research, various advanced algorithms have been applied. Liu et al. proposed a real-time wide-area complex scene object detection framework based on Faster-RCNN. This framework proposes a novel multi-scale feature extraction method that extracts features within receptive fields of different sizes by emphasizing the proportional differences between similar objects, fully capturing the details between objects of different scales in wide-area complex scene object images. Through an adaptive multi-scale feature enhancement module, it effectively distinguishes multi-scale features extracted in hierarchical structures and assigns greater weight to key scale features. Furthermore, researchers are actively exploring computationally more efficient single-level detection algorithms. Jian et al. proposed a self-supervised ocean ship detection method based on Mask R-CNN for small-scale, wide-area complex scene datasets. This method designs a self-supervised task specifically for small target detection, reducing the number of labeled images and improving detection accuracy. Zhou et al. proposed a real-time detection model based on YOLO, SMA-YOLO. This model improves the ability to detect small targets in wide-area complex scenes by designing a bidirectional multi-branch auxiliary feature pyramid network (BIMA-FPN) to fuse high- and low-level features and expand the multi-scale receptive field. Liu innovatively proposed the YOLO-SSP model for target detection in wide-area complex scenes, which introduces a pyramid spatial attention mechanism (PYSAM). It highlights important information in the spatial dimension through multi-level fusion operations, minimizing the impact of complex backgrounds on wide-area complex scene images.

[0004] Overall, while current detection methods for wide-area complex scenes have made significant progress in feature extraction and detection accuracy, challenges remain due to the high frequency of small targets and complex background interference, leading to false positives and false negatives. Furthermore, although some improvements have increased detection accuracy, they have also increased model complexity and computational cost, limiting their effectiveness in real-time applications. Summary of the Invention

[0005] To address the problems existing in the above-mentioned traditional methods, this invention proposes a target detection method, device, equipment, and storage medium for wide-area complex scenes. This method has higher feature sensitivity for dense small targets and irregular targets in wide-area complex scene images, and effectively improves detection accuracy while ensuring the model is lightweight.

[0006] To achieve the above objectives, the embodiments of the present invention adopt the following technical solutions:

[0007] On the one hand, a target detection method for wide-area complex scenes is provided, including the following steps:

[0008] A backbone network is used to extract multi-level features from wide-area complex scene images, resulting in low-level features, mid-level features, and high-level features.

[0009] The EDFPT module is used to encode high-level features to obtain encoded features. The EDFPT module is used to process high-level features using a dual-domain linkage optimization method that combines an efficient discriminative frequency domain feedforward network with polarity-aware linear attention components to obtain encoded features.

[0010] Low-level features, mid-level features, and encoded features are input into the feature fusion module to obtain fused features. The feature fusion module is used to process low-level features, mid-level features, and encoded features using convolution, multi-path feature enhancement, and concatenation operations to obtain fused features. The multi-path feature enhancement module is used to extract and fuse multi-scale and irregular features using a multi-branch parallel architecture, through multi-path feature transformation and StarBlock omnidirectional enhancement mechanism.

[0011] An IoU-aware query selection strategy is used to filter a fixed number of image features from the fused features to obtain an initial query vector.

[0012] The initial query vector is processed by a decoder with an auxiliary prediction head to obtain the target detection results for wide-area complex scene images.

[0013] On the other hand, a target detection device for wide-area complex scenes is also provided, including:

[0014] The feature extraction unit is used to extract multi-level features from wide-area complex scene images using a backbone network, resulting in low-level features, mid-level features, and high-level features.

[0015] The feature encoding unit is used to encode high-level features using the EDFPT module to obtain encoded features; the EDFPT module is used to process high-level features using a dual-domain linkage optimization method that combines an efficient discriminative frequency domain feedforward network with polarity-aware linear attention components to obtain encoded features.

[0016] The feature fusion unit is used to input low-level features, mid-level features, and encoded features into the feature fusion module to obtain fused features. The feature fusion module is used to process low-level features, mid-level features, and encoded features using convolution, multi-path feature enhancement, and concatenation operations to obtain fused features. The multi-path feature enhancement module is used to extract and fuse multi-scale and irregular features using a multi-branch parallel architecture, through multi-path feature transformation and StarBlock omnidirectional enhancement mechanism.

[0017] The initial query vector determination unit is used to select a fixed number of image features from the fused features using an IoU-aware query selection strategy to obtain the initial query vector.

[0018] The target detection result unit is used to process the initial query vector using a decoder with an auxiliary prediction head to obtain the target detection result of the wide-area complex scene image.

[0019] On the other hand, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the above-mentioned target detection methods in wide-area complex scenarios.

[0020] Furthermore, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the steps of any of the above-mentioned target detection methods in wide-area complex scenarios.

[0021] One of the above technical solutions has the following advantages and beneficial effects:

[0022] The aforementioned target detection method, apparatus, device, and storage medium for wide-area complex scenes include the following steps: First, a backbone network is used to extract multi-level features from the wide-area complex scene image, obtaining low-level, mid-level, and high-level features. Second, an EDFPT module is used to encode the high-level features, obtaining encoded features. Third, the low-level, mid-level, and encoded features are input into a feature fusion module to obtain fused features. Fourth, an IoU-aware query selection strategy is used to select a fixed number of image features from the fused features to obtain an initial query vector. Fifth, the initial query vector is processed by a decoder with an auxiliary prediction head to obtain the target detection result for the wide-area complex scene image. This method exhibits higher feature sensitivity for dense small targets and irregularly shaped targets in wide-area complex scene images. While maintaining a lightweight model, it effectively improves detection accuracy and reduces the false negative rate compared to RT-DETR. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application or the conventional technology, the drawings used in the description of the embodiments or the conventional technology will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating a target detection method for a wide-area complex scene in one embodiment;

[0025] Figure 2 This is a schematic diagram of the target detection model structure in a wide-area complex scene in one embodiment;

[0026] Figure 3 Here is an overall framework diagram of the FasterAGCA module in one embodiment;

[0027] Figure 4 This is a general framework diagram of the AGCA module in one embodiment;

[0028] Figure 5 Here is an overall framework diagram of the EDFPT module in one embodiment;

[0029] Figure 6 This is a general framework diagram of the multi-path feature enhancement module in one embodiment;

[0030] Figure 7 Here is a confusion matrix diagram of the method on a SIMD dataset in one embodiment;

[0031] Figure 8 Heatmaps of different models in one embodiment;

[0032] Figure 9 This is a diagram showing the experimental results in one embodiment. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0034] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.

[0035] It should be noted that, in this document, the reference to "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The presentation of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. Those skilled in the art will understand that the embodiments described herein can be combined with other embodiments. The term "and / or" as used herein refers to any combination of one or more of the associated listed items, and all possible combinations, including such combinations.

[0036] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0037] In one embodiment, such as Figure 1 As shown, a target detection method for wide-area complex scenes is provided, which may include the following processing steps 100 to 104:

[0038] Step 100: Use a backbone network to extract multi-level features from wide-area complex scene images to obtain low-level features, mid-level features and high-level features.

[0039] Specifically, the backbone network uses an improved ResNet network, which inputs wide-area complex scene images into the backbone network to extract multi-scale features from the wide-area complex scene images.

[0040] Step 101: Encode the high-level features using the EDFPT module to obtain the encoded features; the EDFPT module is used to process the high-level features using a dual-domain linkage optimization method that combines an efficient discriminative frequency domain feedforward network with polarity-aware linear attention components to obtain the encoded features.

[0041] Specifically, since deep features have richer semantic information and are more suitable for processing by Transformer, the neck network uses a Transformer encoding layer to process only the high-level features output from the last layer of the backbone network.

[0042] To address the limitations of traditional attention mechanisms in handling multi-scale feature differentiation and weak target signal enhancement in wide-area complex scenes where target scale differences are significant and small target features are easily obscured by the background, this application designs an EDFPT (Efficient Discriminative Frequency-Polar Transformer) module. EDFPT integrates an efficient discriminative frequency-domain feedforward network (EDFFN) to enhance the discriminative ability of multi-scale frequency domain features and polar linear attention (PLA) to achieve precise target feature focusing. PLA explicitly separates the polarity interaction of query key pairs and introduces a channel-wise power function with a learnable exponent, enhancing expressive power while maintaining the computational efficiency of linear attention. This dual-component collaborative optimization method effectively distinguishes targets of different scales and highlights small target features obscured by the background, significantly improving the model's localization accuracy and recognition robustness in wide-area complex scene target detection tasks.

[0043] Step 102: Input the low-level features, mid-level features, and coding features into the feature fusion module to obtain the fused features; the feature fusion module is used to process the low-level features, mid-level features, and coding features using convolution, multi-path feature enhancement module, and concatenation operation to obtain the fused features; the multi-path feature enhancement module is used to extract and fuse multi-scale and irregular features using a multi-branch parallel architecture, through multi-path feature transformation and StarBlock omnidirectional enhancement mechanism.

[0044] Specifically, the feature fusion module is used to perform multi-scale feature fusion of low-level features, mid-level features, and coding features to obtain fused features.

[0045] A multi-path feature enhancement module (MFEM) is introduced into the feature fusion (CCFF) module to improve the feature representation capability for irregular targets.

[0046] MFEM employs a multi-path parallel processing architecture. After an initial 1×1 convolution, the input features are split into multiple parallel paths: including a direct convolution transformation path, a StarBlock enhancement path, and a multi-StarBlock cascade path. Different paths respectively complete local feature extraction, deep feature enhancement, and multi-stage feature refinement. Finally, the multi-source features are fused and then 1×1 convolution is performed to complete dimension alignment.

[0047] Step 103: Use an IoU-aware query selection strategy to filter a fixed number of image features from the fused features to obtain an initial query vector.

[0048] Specifically, to address the inconsistency between classification scores and location confidence distributions, an IoU-aware query selection strategy is proposed. This strategy introduces IoU constraints during training to provide higher-quality initial object queries, thereby improving detection performance.

[0049] Step 104: Process the initial query vector using a decoder with an auxiliary prediction head to obtain the target detection results for wide-area complex scene images.

[0050] A wide-area, complex-scene object detection model, abbreviated as WGM-DETR, consists of a backbone network, an EDFPT module, a feature fusion module, an IoU-aware query selection strategy, and a decoder with an auxiliary prediction head. The structure of the wide-area, complex-scene object detection model is as follows: Figure 2 As shown. Figure 2 S3 represents low-level features, S4 represents mid-level features, S5 represents high-level features, and F5 represents coding features.

[0051] The aforementioned target detection method for wide-area complex scenes includes: extracting multi-level features from the wide-area complex scene image using a backbone network to obtain low-level features, mid-level features, and high-level features; encoding the high-level features using an EDFPT module to obtain encoded features; inputting the low-level features, mid-level features, and encoded features into a feature fusion module to obtain fused features; using an IoU-aware query selection strategy to select a fixed number of image features from the fused features to obtain an initial query vector; and processing the initial query vector using a decoder with an auxiliary prediction head to obtain the target detection result for the wide-area complex scene image. This method has higher feature sensitivity for dense small targets and irregularly shaped targets in wide-area complex scene images, effectively improving detection accuracy while maintaining model lightweightness, and reducing the false negative rate compared to RT-DETR.

[0052] In one embodiment, the backbone network is obtained by replacing the BasicBlock module in the ResNet network with the FasterAGCA module; the FasterAGCA module includes PConv, two point convolutional layers, an adaptive graph channel attention module, and residual connections; step 100 includes: processing the wide-area complex scene image sequentially through three CBS modules to obtain convolutional features; wherein the CBS module includes a Conv layer, a BN layer, and a Silu activation function; inputting the convolutional features into the first FasterAGCA module to obtain the first intermediate features; specifically, in the first FasterAGCA module: processing the convolutional features sequentially through PConv, two point convolutional layers, and an adaptive graph channel attention module, and then adding and fusing them with the convolutional features to obtain the first intermediate features; inputting the first intermediate features into the second FasterAGCA module to obtain low-level features; inputting the low-level features into the third FasterAGCA module to obtain mid-level features; inputting the mid-level features into the fourth FasterAGCA module to obtain high-level features.

[0053] Specifically, the backbone network adopts an improved ResNet network; FasterNet is used to optimize BasicBlock, which significantly reduces model complexity and improves inference speed. At the same time, the AGCA mechanism is introduced to effectively suppress background noise interference and enhance the ability to preserve target feature information.

[0054] In the backbone network, the original BasicBlock module is replaced with the lightweight feature extraction module FasterAGCA. This module improves the detection and recognition capabilities of targets while reducing the amount of computation and parameters.

[0055] Due to hardware resource constraints on wide-area complex scene object detection deployment platforms, models need to possess lightweight architecture and low-latency inference capabilities. However, current mainstream wide-area complex scene detection models, such as those based on deep CNNs or Transformers, suffer from high computational redundancy in feature extraction modules and excessively large parameter sizes, leading to a surge in floating-point operations. This makes them unsuitable for resource-constrained wide-area complex scene platforms, directly impacting real-time monitoring response speed. Existing research often reduces model complexity through lightweight modifications, such as depthwise separable convolution (DSConv) and channel pruning techniques. However, these methods often come with the side effect of weakened feature representation capabilities, especially in wide-area complex scene images with dense small targets and complex background features, easily causing the loss of key features and resulting in decreased detection accuracy. Therefore, it is necessary to introduce targeted compensation mechanisms in subsequent network modules to balance lightweight design and detection performance.

[0056] To obtain efficient detection results without relying on high-performance hardware computing devices, Chen et al. designed the efficient and fast network architecture FasterNet. Specifically, they designed a novel partial convolution (PConv) to replace the normal convolutional FasterAGCA module. The FasterAGCA module is as follows: Figure 3 As shown, regular convolution is performed on cp channels of the feature map with input h×w×c, while the remaining channels remain unchanged. Therefore, the floating-point computation of PConv is shown in equation (1):

[0057] (1)

[0058] in, For floating-point operations, c p is the number of channels in the feature map, k is the kernel size, and h and w are the height and width of the feature map, respectively.

[0059] In wide-area complex scene target detection tasks, the acquisition of wide-area complex scene images is affected by various factors such as satellite or UAV shooting altitude, terrain, and weather conditions. Different images exhibit significant differences in target scale and background complexity. Existing feature extraction modules have limitations in processing such wide-area complex scene images involving drastic changes in target scale and complex background interference, struggling to efficiently distinguish targets from the background, especially failing to retain sufficient feature information for small-scale wide-area complex scene targets, resulting in detection performance that falls short of requirements. To address the feature discrimination and extraction problems caused by variable target scale and background noise interference in wide-area complex scenes, Adaptive Graph Channel Attention (AGCA) technology is introduced into the FasterNet module. For example... Figure 4 As shown, the AGCA module can be integrated into the feature map processing flow after each convolutional layer: First, it directly performs nonlocal (NL) operations on the features, avoiding complex pixel-level spatial encoding; then, it converts the feature map dimension into a channel-dimensional representation through global average pooling (GAP), and further enhances the representativeness of the features with the help of a linear embedding function; next, it assigns adaptive weights to the feature vertices of each channel through the adaptive graph convolution module (AGCM); finally, it maps the learned channel weights back to the original feature map. Its lightweight operation reduces computational costs while fully preserving the feature information of each channel, ensuring the inference efficiency of the model in wide-area complex scene object detection tasks. Therefore, this module allows the feature extraction process to focus more on the target and key target regions, possessing stronger targeting and effectiveness in complex wide-area complex scenes containing multi-scale targets, thereby improving the model's detection performance for wide-area complex scene targets. The formula for the AGCA module is as follows:

[0060] (2)

[0061] in This is a global average pooling operation that compresses the feature map in the spatial dimension into a channel-dimensional representation. It is a linear embedding function used to transform the dimensions of features to enhance their representativeness. It is applied twice to features at different stages in the formula. The linear embedding function used in the mapping process. Features that are input into the AGCA module; It is an adaptive graph convolution operation. It is an adaptive learning graph adjacency matrix used to assign adaptive weights to channel feature vertices, enabling information exchange between channels; It is a modified linear unit activation function, used to introduce nonlinearity to enhance feature representation capabilities; the final output is the original feature map multiplied element-wise by the channel attention weights, which strengthens the features of important channels and thus improves the feature capture capability of targets in wide-area complex scenes.

[0062] In one embodiment, the EDFPT module includes: an efficient discriminative frequency domain feedforward network and a polarity-aware linear attention module; step 101 includes: activating high-level features through a dynamic Tanh function and inputting them into the polarity-aware linear attention module to obtain target spatial domain encoded features; adding and fusing the target spatial domain encoded features with the high-level features to obtain intermediate encoded features; activating the intermediate encoded features through a dynamic Tanh function and inputting them into the efficient discriminative frequency domain feedforward network to obtain target frequency domain encoded features; adding and fusing the target frequency domain encoded features with the intermediate encoded features to obtain encoded features.

[0063] Specifically, traditional Feature Attention Interaction (AIFI) modules have significant limitations in wide-area complex scene target detection tasks. On the one hand, their ability to distinguish multi-scale target features is limited, making it difficult to effectively decouple cross-scale features from large buildings to small vehicles in wide-area complex scenes, leading to confusion in feature representations of targets at different scales. On the other hand, AIFI's ability to enhance weak target signals is insufficient, failing to accurately extract features of small targets obscured by complex backgrounds, and it lacks effective utilization of the target's frequency domain characteristics, while the scale differences and saliency features of targets in wide-area complex scenes often exhibit clear pattern differences in the frequency domain. To address these issues, this application proposes an efficient frequency domain polarity discriminator Transformer (EDFPT) module, such as... Figure 5As shown, this EDFPT module significantly improves the model's ability to perceive and distinguish targets in wide-area complex scenes through an innovative combination of an efficient discriminative frequency domain feedforward network (EDFFN) and polar linear attention (PLA): EDFFN focuses on enhancing the discriminativity of multi-scale frequency domain features, accurately capturing different feature targets in multi-scale frequency domain patterns; PLA, by explicitly separating the polar interaction of query key pairs and introducing a channel-wise power function with a learnable exponent, enhances expressive power while maintaining the computational efficiency of linear attention, achieving precise focusing on target features. This dual-component collaborative dual-domain linkage optimization method can effectively distinguish targets of different scales and highlight the features of small targets obscured by the background, significantly improving the model's localization accuracy and recognition robustness in wide-area complex scene target detection tasks.

[0064] First, the features are processed using PLA. PLA maintains the original feature space dimension while explicitly separating the polarity interaction of query key pairs and introducing a channel-wise power function with a learnable exponent, efficiently integrating global contextual information. It also enhances feature representation while maintaining the computational efficiency of linear attention. The processed features are then merged with the original features via residual connections and normalized using LayerNorm. Next, EDFFN further enhances the features by strengthening the discriminative power of multi-scale frequency domain features to adjust the frequency domain features, improving the ability to capture frequency domain patterns specific to targets of different scales and small targets obscured by the background. The forward propagation process of the entire module can be described by the following composite formula:

[0065] (3)

[0066] In this formula, LN represents the input features of the layer, and LN represents the layer normalization operation. PLA stands for Polar Linear Attention, and EDFFN stands for Highly Discriminative Frequency Domain Feedforward Network.

[0067] The PLA module enhances the expressive power of the attention mechanism while maintaining high computational efficiency through explicit polarity interaction separation and linear computation strategies. The module first generates query vectors, key vectors, and value vectors from the input features. Then, it performs feature enhancement through polarity interaction separation, channel-wise power function transformation with a learnable exponent, and finally, linear attention computation. The entire process can be described by the following comprehensive formula:

[0068] (4)

[0069] (5)

[0070] (6)

[0071] in, It is element-wise multiplication. and These are the same-polarity interactive filtering function and the opposite-polarity interactive filtering function, where Q, K, and V are the query, key, and value matrix vectors generated by linear projection, respectively, and X is the feature sequence of the input module. Capture the collaborative pattern of query-key pairs for branches of the same polarity. Capture adversarial modes for heteropolar branches. and It is a learnable gated vector. and It consists of homopolar and heteropolar branches. Through the dual-branch design of "homophonic branches plus heteropolar branches", it solves the core problems of "small target features being obscured by the background" and "large differences in target scale" in wide-area complex scenes while maintaining computational efficiency, thereby improving detection accuracy and robustness.

[0072] EDFFN aims to enhance the discriminative ability of target features in wide-area complex scenes through spatial-frequency domain collaboration. It first extracts multi-scale spatial features and performs nonlinear transformations on the input features using depthwise separable convolutions and point convolutions. Then, the features are rearranged into a local window format to adapt to the "local distribution and multi-scale mixing" characteristics of targets in wide-area complex scenes, facilitating subsequent frequency domain transformation. The rearranged features are converted to the frequency domain using a two-dimensional Fast Fourier Transform (FFT) and modulated using a learnable weight matrix W to selectively enhance or suppress different frequency components. High-frequency components related to small target details and low-frequency components related to large target contours are specifically enhanced, while invalid frequency components from complex backgrounds are suppressed. The frequency-domain modulated features are transformed back to the spatial domain using an Inverse Fourier Transform (IFFT) and finally rearranged back to their original shape. The entire spatial-frequency domain collaborative modulation process can be described by the following composite formula:

[0073] (7)

[0074] in, For the output features, and This represents the feature block expansion and block collapse operations. and Represents two-dimensional Fourier transform and inverse transform. For depthwise separable convolution, For point convolution, the two work together to complete the spatial nonlinear transformation, where W is a learnable frequency domain modulation weight matrix. This represents element-wise multiplication.

[0075] The EDFPT module achieves performance improvements in wide-area complex scene target detection tasks through an innovative combination of PLA and EDFFN. PLA accurately focuses on target spatial features through explicit polarity interaction separation and learnable power function enhancement; EDFFN significantly enhances the ability to capture frequency domain patterns unique to targets in wide-area complex scenes through spatial-frequency domain cross-domain collaboration. Together, they solve the core defects of targets in wide-area complex scenes, namely "large scale differences and small target features being easily obscured by the background".

[0076] In one embodiment, the feature fusion module includes: two convolutional layers, four multi-path feature enhancement modules, and a concatenation operation; step 102 includes: processing the encoded features through the first convolutional layer to obtain a first convolutional feature; upsampling the first convolutional feature and concatenating it with intermediate-level features, and inputting the result into the first multi-path feature enhancement module to obtain a first enhanced feature; processing the first enhanced feature through the second convolutional layer and upsampling it, and concatenating the upsampled result with low-level features and inputting it into the second multi-path feature enhancement module to obtain a second enhanced feature; downsampling the second enhanced feature and inputting it into the third multi-path feature enhancement module to obtain a third enhanced feature; downsampling the third enhanced feature and concatenating the downsampled result with the first convolutional feature and inputting it into the first multi-path feature enhancement module to obtain a fourth enhanced feature; and concatenating the second, third, and fourth enhanced features to obtain a fused feature.

[0077] In one embodiment, such as Figure 6 As shown, the multi-path feature enhancement module adopts a multi-path parallel processing architecture, including: a third convolutional layer, a direct convolutional transformation path, a StarBlock enhancement path, a multi-StarBlock cascaded path, and a fourth convolutional layer. The second enhanced feature is downsampled and input into the third multi-path feature enhancement module to obtain the third enhanced feature. This includes: downsampling the second enhanced feature and processing it through the third convolutional layer to obtain the third convolutional feature; inputting the third convolutional feature into the direct convolutional transformation path to obtain local features; the direct convolutional transformation path is used for local feature extraction through point convolutional layers; inputting the third convolutional feature into the StarBlock enhancement path to obtain depth-enhanced features; the StarBlock enhancement path uses the StarBlock module to perform depth feature enhancement through a combination of depthwise separable convolution and gated attention mechanisms; inputting the third convolutional feature into the multi-StarBlock cascaded path to obtain multiple stage features; the multi-StarBlock cascaded path uses cascaded StarBlock modules for multi-stage feature refinement; concatenating the local features, depth-enhanced features, and multiple stage features and processing them through the fourth convolutional layer to obtain the third enhanced feature.

[0078] In one embodiment, the StarBlock enhancement path includes point convolutional layers and StarBlock modules.

[0079] In one embodiment, a multi-StarBlock cascaded path includes n cascaded StarBlock modules; n is an integer greater than 0; the third convolutional feature is input into the multi-StarBlock cascaded path to obtain multiple stage features, including: channel segmentation of the third convolutional feature to obtain a first segmentation feature and a second segmentation feature; processing the first segmentation feature through a first StarBlock module to obtain a first stage feature; inputting the first stage feature into a second StarBlock module to obtain a second stage feature; and so on, using the feature output by the previous StarBlock module as the input of the next StarBlock module, until the last StarBlock module, to obtain the nth stage feature.

[0080] Specifically, in visual tasks such as target recognition in wide-area complex scenes, traditional feature modules suffer from problems such as single feature extraction dimensions, poor adaptability to multi-scale lesions, and insufficient representation of irregular targets. Their linear architecture limits the flexible expansion of the receptive field, making it difficult to distinguish morphologically similar targets in complex field environments or wide-area complex scenes, thus limiting the model's generalization ability and recognition accuracy. To address these pain points, this paper proposes the MFEM module, which adopts a multi-branch parallel architecture and achieves accurate extraction and fusion of multi-scale and irregular features through multi-path feature transformation and StarBlock enhancement mechanisms.

[0081] The MFEM module employs a multi-path parallel processing architecture. After an initial 1×1 convolution, the input features are split into multiple parallel paths: including a direct convolution transformation path, a StarBlock enhancement path, and a multi-StarBlock cascade path. Different paths respectively perform local feature extraction, deep feature enhancement, and multi-stage feature refinement, ultimately fusing multi-source features and performing 1×1 convolution to achieve dimension alignment. Its overall mathematical expression can be described as follows:

[0082] (8)

[0083] in, For the output features, This represents the direct convolution transformation of the first path. Features are input to the MFEM module. This indicates a single StarBlock enhancement operation for the second path. The characteristic part generated by channel decomposition, This indicates a cascading StarBlock operation. This is the point convolution operation for the final fusion layer, used for feature dimension alignment.

[0084] During the feature transformation process, the StarBlock in the second path employs a combination of depthwise separable convolution and gated interaction. The feature transformation process can be represented as follows:

[0085] (9)

[0086] in, This represents depthwise separable convolution. For two 1×1 convolutions, This represents the ReLU6 activation function. Represents element-wise multiplication. Point convolution that maps to the feature dimensions.

[0087] StarBlock, as the core component of the MFEM module, expands the receptive field through large-kernel depthwise separable convolutions and combines a gated attention mechanism to enhance effective features and suppress redundant information. In wide-area complex scene target detection scenarios, this design can enhance the feature representation capability of irregularly shaped targets such as buildings and ships.

[0088] In summary, the MFEM module significantly improves multi-scale feature capture capabilities and robustness in identifying irregularly shaped targets through the innovative integration of multi-path feature extraction and the StarBlock enhancement mechanism. In rice disease detection tasks, its multi-branch architecture and gating enhancement mechanism significantly improve the model's accuracy in identifying small lesions and morphologically similar diseases, providing an efficient feature extraction paradigm for visual tasks in complex scenarios.

[0089] In a verification embodiment, the hardware and software used were an Intel® Core™ i5-12600K processor with 32 GB of RAM, a GeForce RTX 4060Ti graphics processor with 16 GB of VRAM, PyTorch 11.7 and Python 3.10 as the deep learning model framework, Windows 11 as the operating system, and a cosine annealing learning rate scheduling strategy. To ensure fairness and comparability of model performance, no pre-training weights were used during the training process of any model in all ablation and comparison experiments. During training, a cosine annealing learning rate scheduling strategy was adopted, gradually reducing the learning rate in the later training stages to achieve finer parameter optimization. Online data augmentation strategies were not enabled in any experiments. All other parameters used the default settings of RT-DETR. Detailed parameter settings are shown in Table 1.

[0090] Table 1 Experimental parameters

[0091]

[0092] (1) Dataset

[0093] The NWPU VHR-10 dataset contains 800 high-resolution satellite images cropped from Google Earth and Vaihingen datasets and manually annotated by experts. The dataset is divided into ten categories: Plane (PL), Ship (S), Storage Tank (ST), Baseball Diamond (BD), Tennis Court (TC), Basketball Court (BC), Ground Track Field (GTF), Harbor (HA), Bridge (BR), and Vehicle (V). The dataset includes negative and positive image sets. The negative image set contains 150 images that do not contain any objects of a given object category, while the positive image set contains 650 images; each image contains at least one object of interest.

[0094] The SIMD dataset is a public wide-area complex scene image object detection dataset containing 5000 images with a resolution of 1024×768 pixels and 45,096 instances. It includes a large number of small objects, divided into 15 classes: Car (C), Truck (T), Van (V), Long Vehicle (LV), Bus (B), Airliner (A), Propeller Airplane (PA), Trainer Aircraft (TA), Charter Aircraft (CA), Cargo Aircraft (CA), Other (O), Stair Truck (ST), Pulling Truck (PT), Helicopter (H), and Ship (S).

[0095] (2) Evaluation indicators

[0096] This study uses precision (P), recall (R), and average precision (mAP) to evaluate the model's accuracy in detecting small targets. The calculation methods are detailed in Equations (10) to (13). To thoroughly evaluate the computational efficiency of the model, other metrics are introduced, including the total number of parameters, model size, number of floating-point operations (GFLOPs), and frames per second (FPS).

[0097] (10)

[0098] (11)

[0099] (12)

[0100] (13)

[0101] In these equations, TP represents a correctly predicted true positive, FP represents a false positive that is incorrectly predicted as positive, FN represents a false negative that is incorrectly predicted, and K represents the number of categories.

[0102] (3) Ablation test

[0103] To evaluate the contribution of each improvement to the model, ablation experiments were conducted on the SIMD and NWPU VHR-10 datasets. The following models were defined sequentially: Baseline, Improved Model A (FasterAGCA module), Improved Model B (Efficient Frequency Domain Polarity Discrimination Transformer), Improved Model A+C (Frequency-Aware Local Feature Fusion Module), Improved Model A+B, and Improved Model A+B+C. The changes in six model evaluation metrics were quantitatively explored. The experiments used accuracy (P), recall (R), mean average precision (mAP), model weight file size (Param), frames per second (FPS), and number of parameters as evaluation metrics.

[0104] Under the same experimental conditions, the model experimental results are shown in Table 2. The mAP50, precision, recall, computational cost, and parameters of the baseline model RT-DETR are 78%, 79.9%, 78.2%, 57.3M, and 19.9M, respectively. After adding the FasterAGCA module, the number of parameters decreased by about 13.5%, GFLOPS decreased by about 13.4%, and mAP increased to 0.8%. This is mainly because the FasterAGCA module adopts lightweight multi-scale feature extraction, which significantly reduces the complexity of the model while extracting information at different scales, and enhances the ability to extract features of small targets by introducing the AGCA attention mechanism. The EDFPT module integrates EDFFN and PLA into the AIFI module. The EDFPT module improves localization accuracy and performance through a dual-component collaborative dual-domain linkage optimization method. After adding EDFPT, the model's recall rate significantly improved by 2.0%, reducing missed detections in wide-area complex scene image target detection, while mAP50 also increased. The MFEM module enhances the distinction between targets and backgrounds in complex scenes through multi-path feature enhancement, reducing false detections. Ultimately, the combination of these modules enables WGM-DETR to achieve optimal detection performance, with an mAP50 of 81.5%, a 3.5% improvement over the baseline model, significantly enhancing target detection performance in wide-area complex scenes.

[0105] Table 2 Ablation experiments on the SIMD dataset

[0106]

[0107] Similar experimental results were obtained on the NWPU VHR-10 dataset. As shown in Table 3, the network integrating multiple modules performs significantly better than the network integrating only a single module. The synergistic effect of the three modules allows the model to progressively align and refine contextual information from multiple layers of the network, while efficiently fusing features from both deep and shallow layers, and significantly suppressing the interference of complex backgrounds on target features, ultimately achieving more accurate target regression and category judgment. The experiments also verified the good compatibility between the proposed modules; when all modules are fully integrated, the model performance reaches its optimal level, with an mAP50 index of 92.1%.

[0108] Table 3 Ablation experiments on the NWPU VHR-10 dataset

[0109]

[0110] (4) Comparative experiment

[0111] like Figure 7As shown, the confusion matrix of the upgraded model is illustrated, with the horizontal axis representing the actual values ​​and the vertical axis representing the predicted values. This matrix provides an intuitive summary, reflecting the model's performance by comparing the actual results with the predictions. The graph shows that most predictions are consistent with the actual values, indicating that the model has strong predictive power. Under the same experimental conditions, the detection performance of the improved WGM-DETR is significantly different from that of the original RT-DETR in all aspects. To further verify the superiority of WGM-DETR in wide-area complex scene image target detection tasks, this embodiment compares it with classic models on the SIMD dataset, considering both the number of parameters and accuracy metrics. The selected comparison algorithms include Faster-RCNN, EfficientDet, YOLOv3, YOLOv5, YOLOv8, YOLOv10, YOLOv11, and Deformable-DETR. The experimental results are shown in Table 4.

[0112] Comparative experiments were conducted on the SIMD dataset using WGM-DETR and mainstream object detection models. Compared to the classic two-stage model Faster R-CNN, WGM-DETR significantly reduces the number of parameters while substantially improving mAP50. Compared to the lightweight model EfficientDet, WGM-DETR demonstrates a clear advantage in detecting complex objects in wide-area, complex scenes, exhibiting superior accuracy. Compared to the YOLO series models, while YOLOv3 shows some improvement in accuracy, its 103.7M parameters limit its practicality in real-world scenarios. Models like YOLOv5m and YOLOv8m show good accuracy, but their parameter counts are higher than WGM-DETR. YOLOv10m, with 16.4M parameters, still slightly lags behind WGM-DETR in accuracy. Furthermore, Deformable-DETR, with 40.0M parameters, also falls short of WGM-DETR in accuracy. In summary, WGM-DETR demonstrates outstanding performance in balancing parameter quantity and detection accuracy. With only 17.2M parameters, it achieves mAP50 and mAP50-95 of 78.0% and 65.8% respectively on the SIMD dataset, fully validating its superiority in wide-area complex scene image target detection tasks. It achieves a high-efficiency balance between model lightweighting and detection accuracy, making it more suitable for deployment requirements in real-world wide-area complex scenes.

[0113] Table 4 Performance comparison of different network models on the SIMD dataset

[0114]

[0115] (5) Generalization experiment

[0116] To verify the model's adaptability and generalization ability in wide-area complex scene object detection, this embodiment conducted experiments on the RSOD dataset and performed comparative experiments on the SIMD, NWPU VHR-10, and RSOD datasets. The experimental results are shown in Table 5. WGM-DETR achieved significant improvements in P, R, and mAP on the SIMD and NWPU VHR-10 datasets. Although the recall rate decreased slightly on the RSOD dataset, the significant improvement in model accuracy ensured that it maintained high detection performance while reducing the false positive rate, demonstrating the strong advantages of WGM-DETR in wide-area complex scene object detection tasks.

[0117] Table 5 Comparison results on different datasets

[0118]

[0119] (6) Analysis of experimental results

[0120] (a) Model heatmap:

[0121] To visually demonstrate the detection performance of WGM-DETR, this embodiment uses visualization technology to generate a heatmap as shown below. Figure 8 As shown, where Figure 8 The first column is the original image, the second column is the RT-DETR model heatmap, and the third column is the WGM-DETR model heatmap. Pixels with higher gradients in the feature map are represented by darker red in the heatmap, indicating more attention to these areas, while pixels with lower gradients are depicted by darker blue, indicating reduced sensitivity to these areas. In large target scenarios, the WGM-DETR model heatmap in the third column shows each aircraft completely covered by a more concentrated and distinct red heatmap. Compared to intermediate models, it demonstrates higher overall attention to large targets and more comprehensive feature capture, accurately focusing on key areas of large targets and enhancing the stability of large target recognition. In dense small target scenarios, numerous densely distributed small targets form clear and relatively independent red attention areas, effectively distinguishing dense small targets and preventing them from being submerged due to their dense distribution, thus significantly improving the detection accuracy of dense small targets. In occluded target scenarios, WGM-DETR allows occluded targets to present clearer red heatmaps. Compared to intermediate models, it is better at capturing occlusion factors and mining effective target features, significantly enhancing the perception and detection capabilities of occluded targets and improving detection performance in complex occluded scenarios.

[0122] (b) Detection results comparison: In order to more vividly evaluate the detection performance of WGM-DETR in dense small targets, multi-scale targets and complex backgrounds of wide-area complex scene images, WGM-DETR and the baseline model RT-DETR were compared on the SIMD dataset. Three typical images with the characteristics of dense small boats, multi-scale vehicles and multi-class mixture with complex backgrounds were selected from the dataset. In the first row of the boat scene, although RT-DETR can detect most boats and provide classification confidence scores, WGM-DETR performs better in terms of the number of boats detected and confidence scores, and is more accurate in identifying densely packed small boats. In the second row of the vehicle scene, WGM-DETR is superior to RT-DETR in terms of the completeness and confidence of vehicle detection, and has higher accuracy in the localization and classification of vehicles at multiple scales. In the third row of the multi-class mixed scene, RT-DETR has some issues with missed detections and misclassification, while WGM-DETR almost accurately identifies all targets, whether vehicles or boats. It maintains high detection accuracy and class accuracy even in complex backgrounds, demonstrating strong adaptability to multi-class, complex backgrounds and wide-area complex scenes. In summary, WGM-DETR demonstrates superior detection performance compared to RT-DETR in dense small target, multi-scale target, and complex background scenes in the SIMD dataset. It improves in the number of target detections, classification confidence, and scene adaptability, reflecting its strong feature extraction and accurate localization capabilities in small target detection tasks in wide-area complex scene images.

[0123] like Figure 9 As shown, Figure 9 The first column shows the original image, the second column shows the experimental results of the RT-DETR model, and the third column shows the experimental results of the WGM-DETR model. Experimental results on the SIMD and NWPU VHR-10 datasets validate the feasibility and superiority of this method. Without pre-training, WGM-DETR achieves an mAP of 81.5% on the SIMD dataset with only 17.2M parameters, and an mAP of 92.1% on the NWPU VHR-10 dataset, both significantly higher than the original RT-DETR. Especially in tasks such as dense small target detection and irregular target recognition in complex wide-area scenes, it reduces the false negative rate by approximately 3.2% compared to RT-DETR, fully demonstrating its performance advantage.

[0124] This application proposes the WGM-DETR object detection model for wide-area complex scenes. The WGM-DETR algorithm optimizes the backbone feature extraction layer and Transformer encoding / decoding structure of RT-DETR, and designs a lightweight multi-scale feature extraction module (FasterAGCA), an efficient discriminative frequency domain polarity Transformer (EDFPT) module, and a multi-path feature enhancement module (MFEM). It constructs a dedicated feature extraction and detection network for dense small targets, irregularly shaped targets, and complex backgrounds in wide-area complex scenes. This model exhibits higher feature sensitivity for dense small targets and irregularly shaped targets in wide-area complex scene images, effectively improving detection accuracy while maintaining a lightweight model. Experimental results on the SIMD and NWPU VHR-10 datasets verify the feasibility and superiority of this method. Without pre-training, WGM-DETR achieves an mAP of 81.5% on the SIMD dataset with only 17.2M parameters, and an mAP of 92.1% on the NWPU VHR-10 dataset, both significantly higher than the original RT-DETR. Especially in tasks such as dense small target detection and irregular target recognition in complex wide-area scenarios, the false negative rate is reduced by about 3.2% compared to RT-DETR, which fully demonstrates the performance advantage.

[0125] It should be understood that, although the above Figure 1 The steps are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise explicitly stated in this document, there is no strict order in which these steps are executed; they can be performed in other orders. Furthermore, the above... Figure 1 At least some of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0126] In one embodiment, a target detection device for wide-area complex scenes is also provided, comprising:

[0127] The feature extraction unit is used to extract multi-level features from wide-area complex scene images using a backbone network, resulting in low-level features, mid-level features, and high-level features.

[0128] The feature encoding unit is used to encode high-level features using the EDFPT module to obtain encoded features; the EDFPT module is used to process high-level features using a dual-domain linkage optimization method that combines an efficient discriminative frequency domain feedforward network with polarity-aware linear attention components to obtain encoded features.

[0129] The feature fusion unit is used to input low-level features, mid-level features, and encoded features into the feature fusion module to obtain fused features. The feature fusion module is used to process low-level features, mid-level features, and encoded features using convolution, multi-path feature enhancement, and concatenation operations to obtain fused features. The multi-path feature enhancement module is used to extract and fuse multi-scale and irregular features using a multi-branch parallel architecture, through multi-path feature transformation and StarBlock omnidirectional enhancement mechanism.

[0130] The initial query vector determination unit is used to select a fixed number of image features from the fused features using an IoU-aware query selection strategy to obtain the initial query vector.

[0131] The target detection result unit is used to process the initial query vector using a decoder with an auxiliary prediction head to obtain the target detection result of the wide-area complex scene image.

[0132] In one embodiment, the backbone network is obtained by replacing the BasicBlock module in the ResNet network with the FasterAGCA module. The FasterAGCA module includes PConv, two point convolutional layers, an adaptive graph channel attention module, and residual connections. The feature extraction unit is also used to process the wide-area complex scene image sequentially through three CBS modules to obtain convolutional features. The CBS module includes a Conv layer, a BN layer, and a Silu activation function. The convolutional features are input into the first FasterAGCA module to obtain the first intermediate features. Specifically, in the first FasterAGCA module: the convolutional features are processed sequentially through PConv, two point convolutional layers, and an adaptive graph channel attention module, and then added and fused with the convolutional features to obtain the first intermediate features. The first intermediate features are input into the second FasterAGCA module to obtain low-level features. The low-level features are input into the third FasterAGCA module to obtain mid-level features. The mid-level features are input into the fourth FasterAGCA module to obtain high-level features.

[0133] In one embodiment, the EDFPT module includes: an efficient discriminative frequency domain feedforward network and a polarity-aware linear attention module; a feature encoding unit, further configured to input high-level features after activation by a dynamic Tanh function into the polarity-aware linear attention module to obtain target spatial domain encoded features; to add and fuse the target spatial domain encoded features with the high-level features to obtain intermediate encoded features; to activate the intermediate encoded features by a dynamic Tanh function and input them into the efficient discriminative frequency domain feedforward network to obtain target frequency domain encoded features; and to add and fuse the target frequency domain encoded features with the intermediate encoded features to obtain encoded features.

[0134] In one embodiment, the feature fusion module includes: two convolutional layers, four multi-path feature enhancement modules, and a concatenation operation; the feature fusion unit is further configured to process the encoded features through the first convolutional layer to obtain a first convolutional feature; upsample the first convolutional feature and concatenate it with intermediate-level features, inputting the result into the first multi-path feature enhancement module to obtain a first enhanced feature; upsample the first enhanced feature through the second convolutional layer, concatenate the upsampled result with low-level features, and input it into the second multi-path feature enhancement module to obtain a second enhanced feature; downsample the second enhanced feature and input it into the third multi-path feature enhancement module to obtain a third enhanced feature; downsample the third enhanced feature and concatenate the downsampled result with the first convolutional feature, inputting it into the first multi-path feature enhancement module to obtain a fourth enhanced feature; and concatenate the second, third, and fourth enhanced features to obtain a fused feature.

[0135] In one embodiment, the multi-path feature enhancement module adopts a multi-path parallel processing architecture, including: a third convolutional layer, a direct convolutional transformation path, a StarBlock enhancement path, a multi-StarBlock cascaded path, and a fourth convolutional layer; a feature fusion unit is further used to downsample the second enhanced feature and process it through the third convolutional layer to obtain the third convolutional feature; input the third convolutional feature into the direct convolutional transformation path to obtain local features; the direct convolutional transformation path is used to extract local features through point convolutional layers; input the third convolutional feature into the StarBlock enhancement path to obtain depth-enhanced features; the StarBlock enhancement path is used to enhance depth features using a combination of depthwise separable convolution and gated attention mechanisms with the StarBlock module; input the third convolutional feature into the multi-StarBlock cascaded path to obtain multiple stage features; the multi-StarBlock cascaded path is used to refine multi-stage features using cascaded StarBlock modules; and concatenate the local features, depth-enhanced features, and multiple stage features and process them through the fourth convolutional layer to obtain the third enhanced feature.

[0136] In one embodiment, the StarBlock enhancement path in the feature fusion unit includes a point convolutional layer and a StarBlock module.

[0137] In one embodiment, the multi-StarBlock cascaded path includes n cascaded StarBlock modules; n is an integer greater than 0; the feature fusion unit is further used to perform channel segmentation on the third convolutional feature to obtain a first segmented feature and a second segmented feature; after the first segmented feature is processed by the first StarBlock module, a first-stage feature is obtained; the first-stage feature is input into the second StarBlock module to obtain the second-stage feature; and so on, using the feature output by the previous StarBlock module as the input of the next StarBlock module, until the last StarBlock module, to obtain the nth-stage feature.

[0138] It is understood that for detailed explanations of the target detection device in wide-area complex scenes, please refer to the corresponding explanations of the various embodiments of the target detection method in wide-area complex scenes above, and will not be repeated here. Each module in the aforementioned target detection device in wide-area complex scenes can be implemented entirely or partially through software, hardware, or a combination thereof. Each module can be embedded in hardware or independently of a device with data processing capabilities, or stored in software in the memory of the aforementioned device, so that the processor can call and execute the operations corresponding to each module. The aforementioned device can be, but is not limited to, various types of data processing computer devices already existing in the art.

[0139] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0140] It is understood that, in addition to the memory and processor mentioned above, the computer equipment described above also includes other hardware and software components not listed in this specification. The specific components can be determined according to the model of the image processing computer in different application scenarios, and will not be listed and described in detail in this specification.

[0141] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the steps or sub-steps added in the various embodiments of the target detection method in the wide-area complex scene described above.

[0142] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), memory bus DRAM (RDRAM), and interface DRAM (DRDRAM), etc.

[0143] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0144] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of protection of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and all such modifications and improvements fall within the scope of protection of this application.

Claims

1. A target detection method in a wide area complex scene, characterized in that, Including the following steps: A backbone network is used to extract multi-level features from wide-area complex scene images, resulting in low-level features, mid-level features, and high-level features. The backbone network is obtained by replacing the BasicBlock module in the ResNet network with the FasterAGCA module; the FasterAGCA module includes PConv, two point convolutional layers, an adaptive graph channel attention module, and residual connections. The high-level features are encoded using the EDFPT module to obtain encoded features; Specifically, the process includes: activating the high-level features using a dynamic Tanh function and then inputting them into a polarity-aware linear attention module to obtain target spatial domain encoded features; adding and fusing the target spatial domain encoded features with the high-level features to obtain intermediate encoded features; activating the intermediate encoded features using a dynamic Tanh function and then inputting them into an efficient discriminative frequency domain feedforward network to obtain target frequency domain encoded features; and adding and fusing the target frequency domain encoded features with the intermediate encoded features to obtain encoded features. The EDFPT module is used to process the high-level features using a dual-domain linkage optimization method that combines an efficient discriminative frequency domain feedforward network and a polarity-aware linear attention module to obtain encoded features. The EDFPT module includes an efficient discriminative frequency domain feedforward network and a polarity-aware linear attention module. The low-level features, mid-level features, and encoded features are input into the feature fusion module to obtain fused features. The feature fusion module is used to process the low-level features, mid-level features, and encoded features using convolution, multi-path feature enhancement, and concatenation operations to obtain fused features. The feature fusion module includes: two convolutional layers, four multi-path feature enhancement modules, and one concatenation operation. The multi-path feature enhancement modules are used to extract and fuse multi-scale and irregular features using a multi-branch parallel architecture, through multi-path feature transformation and StarBlock omnidirectional enhancement mechanism. The multi-path feature enhancement module adopts a multi-path parallel processing architecture, including: a third convolutional layer, a direct convolutional transformation path, a StarBlock enhancement path, multiple StarBlock cascaded paths, and a fourth convolutional layer. The StarBlock enhancement path is used to perform deep feature enhancement using the StarBlock module through a combination of depthwise separable convolution and gated attention mechanism. An IoU-aware query selection strategy is used to filter a fixed number of image features from the fused features to obtain an initial query vector; The initial query vector is processed by a decoder with an auxiliary prediction head to obtain the target detection results for a wide-area complex scene image.

2. The target detection method in a wide-area complex scene according to claim 1, characterized in that, A backbone network is used to extract multi-level features from wide-area complex scene images, resulting in low-level features, mid-level features, and high-level features, including: The wide-area complex scene image is processed sequentially through three CBS modules to obtain convolutional features; the CBS module includes a Conv layer, a BN layer, and a Silu activation function; The convolutional features are input into the first FasterAGCA module to obtain the first intermediate features. Specifically, in the first FasterAGCA module, the convolutional features are processed sequentially through PConv, two point convolutional layers, and an adaptive graph channel attention module, and then added and fused with the convolutional features to obtain the first intermediate features. The first intermediate feature is input into the second FasterAGCA module to obtain the low-level features; The low-level features are input into the third FasterAGCA module to obtain the mid-level features; The intermediate-level features are input into the fourth FasterAGCA module to obtain high-level features.

3. The target detection method in a wide-area complex scene according to claim 1, characterized in that, The low-level features, the mid-level features, and the encoded features are input into the feature fusion module to obtain fused features, including: The encoded features are processed through the first convolutional layer to obtain the first convolutional features; The first convolutional feature is upsampled and then concatenated with the intermediate-order feature. The result is then input into the first multi-path feature enhancement module to obtain the first enhanced feature. The first enhanced feature is processed by the second convolutional layer and then upsampled. The upsampled result is concatenated with the low-level feature and then input into the second multi-path feature enhancement module to obtain the second enhanced feature. The second enhanced feature is downsampled and then input into the third multi-path feature enhancement module to obtain the third enhanced feature; After downsampling the third enhanced feature, the downsampling result is concatenated with the first convolutional feature and then input into the first multi-path feature enhancement module to obtain the fourth enhanced feature; The second enhancement feature, the third enhancement feature, and the fourth enhancement feature are concatenated to obtain the fused feature.

4. The target detection method in a wide-area complex scene according to claim 3, characterized in that, The second enhanced feature is downsampled and then input into the third multi-path feature enhancement module to obtain the third enhanced feature, which includes: After downsampling the second enhanced feature, it is processed by the third convolutional layer to obtain the third convolutional feature; The third convolutional feature is input into the direct convolutional transformation path to obtain local features; the direct convolutional transformation path is used to extract local features through point convolutional layers. The third convolutional feature is input into the StarBlock enhancement path to obtain the depth enhancement feature; the StarBlock enhancement path is used to enhance the depth feature by combining depthwise separable convolution and gated attention mechanism using the StarBlock module. The third convolutional feature is input into the multi-StarBlock cascaded path to obtain multiple stage features; the multi-StarBlock cascaded path is used to refine the multi-stage features using cascaded StarBlock modules; The local features, depth enhancement features, and features from multiple stages are concatenated and then processed through a fourth convolutional layer to obtain the third enhancement feature.

5. The target detection method in a wide-area complex scene according to claim 4, characterized in that, The StarBlock enhancement path includes point convolutional layers and the StarBlock module.

6. The target detection method in a wide-area complex scene according to claim 4, characterized in that, The multi-StarBlock cascade path includes n cascaded StarBlock modules; n is an integer greater than 0; The third convolutional feature is input into the multi-StarBlock cascaded path to obtain multiple stage features, including: The third convolutional feature is segmented by channels to obtain the first segmentation feature and the second segmentation feature; After the first segmentation feature is processed by the first StarBlock module, the first stage feature is obtained; The first-stage features are input into the second StarBlock module to obtain the second-stage features; This process continues, using the features output by the previous StarBlock module as the input to the next StarBlock module, until the last StarBlock module is reached, thus obtaining the nth stage features.

7. A target detection device for wide-area complex scenes, characterized in that, include: The feature extraction unit is used to perform multi-level feature extraction on wide-area complex scene images using a backbone network to obtain low-level features, mid-level features, and high-level features. The backbone network is obtained by replacing the BasicBlock module in the ResNet network with the FasterAGCA module. The FasterAGCA module includes PConv, two point convolutional layers, an adaptive graph channel attention module, and residual connections. A feature encoding unit is used to encode the high-level features using an EDFPT module to obtain encoded features. Specifically, this includes: activating the high-level features using a dynamic Tanh function and then inputting them into a polarity-aware linear attention module to obtain target spatial domain encoded features; adding and fusing the target spatial domain encoded features with the high-level features to obtain intermediate encoded features; activating the intermediate encoded features using a dynamic Tanh function and then inputting them into an efficient discriminative frequency domain feedforward network to obtain target frequency domain encoded features; and adding and fusing the target frequency domain encoded features with the intermediate encoded features to obtain encoded features. The EDFPT module is used to process the high-level features using a dual-domain linkage optimization method that combines an efficient discriminative frequency domain feedforward network and a polarity-aware linear attention module to obtain encoded features. The EDFPT module includes an efficient discriminative frequency domain feedforward network and a polarity-aware linear attention module. A feature fusion unit is used to input the low-level features, the mid-level features, and the encoded features into the feature fusion module to obtain fused features. The feature fusion module is used to process the low-level features, mid-level features, and encoded features using convolution, multi-path feature enhancement modules, and concatenation operations to obtain fused features. The feature fusion module includes two convolutional layers, four multi-path feature enhancement modules, and one concatenation operation. The multi-path feature enhancement modules are used to extract and fuse multi-scale and irregular features using a multi-branch parallel architecture, through multi-path feature transformation and StarBlock omnidirectional enhancement mechanisms. The multi-path feature enhancement modules adopt a multi-path parallel processing architecture, including a third convolutional layer, a direct convolutional transformation path, a StarBlock enhancement path, multiple StarBlock cascaded paths, and a fourth convolutional layer. The StarBlock enhancement path uses the StarBlock module to perform deep feature enhancement through a combination of depthwise separable convolution and gated attention mechanisms. An initial query vector determination unit is used to filter a fixed number of image features from the fused features using an IoU-aware query selection strategy to obtain an initial query vector; The target detection result unit is used to process the initial query vector using a decoder with an auxiliary prediction head to obtain the target detection result of the wide-area complex scene image.

8. A computer device, comprising a memory and a processor, characterized in that, The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the target detection method in a wide-area complex scene as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the target detection method in any one of claims 1 to 6 for wide-area complex scenes.

Images