Target detection method in improved DEIM field complex environment

Abstract

The application discloses a target detection method in a complex field environment based on an improved DEIM, and the method introduces an ODConv module to learn parallel attention of four dimensions of a convolution kernel space, dynamically modulate the convolution kernel, capture subtle features of a target, and effectively improve the distinguishing ability for similar interference objects; a wConv-C3k2 module is constructed, a spatial density function is introduced to dynamically adjust the weight of the convolution kernel, which enables the model to preferentially process the core features of the target, weaken the interference of background noise, and improve the quality of spatial feature extraction; an MDFP module is introduced, semantic, texture and position features are extracted through three complementary branches respectively, small target details are reserved without weighting, key feature loss in the downsampling process is reduced, small target detection capability is significantly enhanced, and the model calculation amount and parameter amount are slightly reduced.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, and in particular to a target detection method based on an improved DEIM (Deep Detection Information Model) for complex field environments. Background Technology

[0002] In wilderness environments such as forests, mountains, and undeveloped areas, the number of people participating in activities such as hiking, exploration, and camping continues to grow. Due to factors such as the type of activity, environmental conditions, and insufficient physical and mental preparation, various safety accidents such as getting lost, injuries, and falls occur frequently, leading to a growing demand for search and rescue operations and assistance to those in distress. Drones, with their advantages of high flexibility, wide coverage, and the elimination of the need for close-range personnel, have become an indispensable technological tool in search and rescue missions. With the development of artificial intelligence technology, deep learning-based target detection technology can assist drones in automatically identifying human targets in images and assessing their approximate condition, providing precise data support for rescue teams to formulate search and rescue plans and allocate rescue resources.

[0003] Outdoor scenes present challenges such as vegetation obstruction, drastic changes in lighting, and complex background textures, which can lead to insignificant differences between similar interfering objects and target features, resulting in false detections by the model. High background noise can also obscure target features, causing missed detections. High-altitude drone photography results in a very low percentage of target pixels, leading to a lack of feature information for small targets. Furthermore, traditional downsampling operations can easily lose key details such as the spatial location and texture edges of small targets, resulting in insufficient accuracy in small target detection.

[0004] Among existing object detection algorithms, the DEIM (DETR with Improved Matching) network exhibits good convergence efficiency and detection accuracy in real-time object detection due to its introduction of a dense one-to-one matching strategy and a matching-aware loss mechanism. Furthermore, the model is highly lightweight and adaptable to UAV edge computing platforms. However, when the benchmark DEIM network is applied to complex outdoor environments, it still suffers from the aforementioned problems of false detection, missed detection, and low accuracy in detecting small objects: its HG-Block module in the backbone network uses static convolutional kernels, focusing only on local receptive field feature matching, and cannot adjust the response by combining global context or feature dependencies of the input image. In addition, static convolution is limited by fixed sampling positions, making it difficult to flexibly adjust the receptive field to adapt to the actual target shape, easily leading to false detections due to insufficient feature discrimination between similar objects and the real target; the RepNCSPELAN4 processing unit in the encoder network uses a standard convolutional method that assigns the same weight to all pixels in the receptive field, resulting in high sensitivity to illumination changes and background noise; traditional convolutional downsampling often blurs semantic boundaries, discards key texture details of small targets, and leads to excessive smoothing of local features. Summary of the Invention

[0005] Therefore, it is necessary to propose an improved DEIM-based target detection method for complex field environments to address the above problems.

[0006] An improved target detection method based on DEIM in complex field environments, the method comprising:

[0007] Images are obtained from the SARD dataset. The obtained images are subjected to data augmentation operations including Mosaic stitching, random photometric distortion, random scaling, random IoU cropping, random horizontal flipping, and Mixup. The first input features are obtained after preprocessing the data-augmented images.

[0008] The backbone network receives the first input feature and outputs the P2 layer features, P3 layer features and P4 layer features to the encoder network;

[0009] The encoder network performs channel dimensionality reduction and projection processing on the features of layer P2, layer P3, and layer P4 through 1×1 convolutional layers to obtain the input features of layer P2, layer P3, and layer P4, respectively. These input features are then fed into the FPN path within the encoder network.

[0010] The P4 layer input features are enhanced by the AIFI module in the FPN path to obtain the P4 layer enhanced features; and based on the P4 layer enhanced features, the second fusion feature, the third fusion feature, and the fourth fusion feature are obtained and input into the decoder network.

[0011] The second, third, and fourth fusion features are received as multi-scale fusion features, and these multi-scale fusion features are used as key-value inputs for cross-attention to obtain multi-scale key-value features. An initialized learnable target query vector is used as the initial query input, and during the training phase, noisy real-labeled auxiliary queries are introduced to form a decoded query set. The decoded query set is input to the self-attention module of the decoding layer for information interaction processing between queries to obtain query interaction features. A prediction result is determined based on the query interaction features and the multi-scale key-value features. The prediction result represents natural human postures such as running, walking, standing, sitting, and lying down.

[0012] In one embodiment, the backbone network includes: a StemBlock layer and four BasicStage layers, wherein the StemBlock layer is a feature extraction module; the backbone network receives the first input features and outputs P2 layer features, P3 layer features, and P4 layer features to the encoder network, including:

[0013] The first input features are initially downsampled by a 3×3 convolution with a stride of 2 in the feature extraction module to obtain the initial features;

[0014] The initial features are input into two branches. The initial features are subjected to max pooling in the first path to obtain salient features. The initial features are then subjected to two 2×2 convolutions in the second path to perform channel-dimensional feature transformation to obtain transformed features. The salient features and transformed features are then concatenated and fused to obtain fused features. The fused features are then downsampled and extracted again by a 3×3 convolution with a stride of 2 to obtain extracted features. The extracted features are then subjected to a 1×1 convolution for channel adjustment to obtain basic features.

[0015] The basic features are input into the first BasicStage layer. The multi-layer ODConv module in the HG-Block module of the first BasicStage layer extracts features layer by layer to obtain multiple first-layer features. The multiple first-layer features are then concatenated with the basic features to obtain multi-scale fusion features. The multi-scale fusion features are input into the SE compression-excited attention module in the first BasicStage layer for channel-dimensional feature recalibration, and the enhanced features are output to the second BasicStage layer.

[0016] In the second BasicStage layer, the input enhancement features are first downsampled by a 3×3 depthwise separable convolution with a stride of 2 to obtain the first downsampled features. The first downsampled features are then extracted layer by layer by the multi-layer ODConv module in the HG-Block module of the second BasicStage layer to obtain multiple second-layer features. The multiple second-layer features are then concatenated with the first downsampled features and aggregated by the SE compression excitation attention module to obtain the P2 layer features. The P2 layer features are then output to the third BasicStage layer of the encoder network.

[0017] In the third BasicStage layer, the P2 layer features are first downsampled by a 3×3 depthwise separable convolution with a stride of 2 to obtain the second downsampled features. The second downsampled features are then extracted layer by layer by the multi-layer ODConv module in the HG-Block module of the third BasicStage layer to obtain multiple third layer features. The multiple third layer features are then concatenated with the second downsampled features and aggregated by the SE compression and excitation attention module to obtain the P3 layer features. The P3 layer features are then output to the encoder network and the fourth BasicStage layer.

[0018] In the fourth BasicStage layer, the P3 layer features are first downsampled using a 3×3 depthwise separable convolution with a stride of 2 to obtain the third downsampled features. The third downsampled features are then extracted layer by layer by the multi-layer ODConv module within the HG-Block module in the fourth BasicStage layer to obtain multiple fourth layer features. These multiple fourth layer features are then concatenated with the third downsampled features and aggregated by the SE compression-excited attention module to obtain the P4 layer features. The P4 layer features are then output to the encoder network.

[0019] In one embodiment, the execution process of the multi-layer ODConv module is as follows:

[0020] The second input feature is subjected to global average pooling processing by an adaptive global average pooling layer to obtain a global pooling feature; the global pooling feature is then subjected to channel dimensionality reduction processing by a 1×1 convolutional layer to obtain a dimensionality-reduced feature; the dimensionality-reduced feature is then subjected to normalization processing and activation processing by a layer normalization layer and a ReLU activation layer in sequence to obtain a channel compression feature; the second input feature is a basic feature, a first downsampled feature, a second downsampled feature, or a third downsampled feature;

[0021] The channel attention weights are obtained by performing fully connected mapping and Sigmoid activation on the channel compressed features based on the channel attention branches.

[0022] The channel compression features are fully connected and Sigmoid activation processed according to the filter attention branch to obtain the filter attention weights.

[0023] Based on the spatial attention branch, the channel compression features are fully connected and Sigmoid activation is applied to obtain the spatial attention weights.

[0024] Based on the convolution kernel attention branch, the channel compression features are fully connected and Softmax normalized to obtain the convolution kernel attention weights.

[0025] The second input feature is multiplied element-wise according to the channel attention weights to obtain the channel-weighted feature;

[0026] Based on the spatial attention weights and the convolutional kernel attention weights, multiple sets of learnable convolutional kernel weight parameters are multiplied element-wise to obtain weighted multi-kernel weights.

[0027] The weighted multi-kernel weights are weighted and aggregated along the convolution kernel dimension to obtain aggregated convolution kernel weights.

[0028] The channel-weighted features are subjected to two-dimensional convolution processing based on the aggregated convolution kernel weights to obtain the initial convolution output features;

[0029] The initial convolutional output features are multiplied element-wise according to the filtering attention weights to obtain a first output feature, wherein the first output feature is a plurality of first-layer features, a plurality of second-layer features, a plurality of third-layer features, or a plurality of fourth-layer features.

[0030] The first output feature is achieved through the following expression:

[0031]

[0032] in, Indicates the second input feature, Indicates the first output feature. For multiple sets of learnable convolutional kernel weight parameters, , , as well as These represent the spatial attention weights, channel attention weights, filter attention weights, and convolutional kernel attention weights, respectively.

[0033] In one embodiment, obtaining the second fusion feature, the third fusion feature, and the fourth fusion feature based on the enhanced features of the P4 layer, and inputting them into the decoder network includes:

[0034] The enhanced features of layer P4 are transmitted in two paths. One path is transmitted to the PAN path, and the other path is upsampled and then concatenated with the input features of layer P3 along the channel dimension to obtain the first concatenated feature.

[0035] The first concatenated feature is input into the wConv-C3k2 module for feature fusion to obtain the first fused feature; the first fused feature is then subjected to a convolution operation to obtain the first convolutional feature;

[0036] The first convolutional feature is passed in two paths. One path is passed to the PAN path, and the other path is upsampled and then concatenated with the input feature of layer P2 along the channel dimension to obtain the second concatenated feature.

[0037] The second concatenated feature is input into the wConv-C3k2 module for feature fusion to obtain the second fused feature; the second fused feature is passed to the decoder network in one path and to the PAN path in the other path;

[0038] In the PAN path, the second fusion feature is downsampled by the MDFP module to obtain the fourth downsampled feature. The fourth downsampled feature is then concatenated with the P4 layer enhancement feature at the channel dimension to obtain the third concatenated feature.

[0039] The third concatenated feature is input into the wConv-C3k2 module for feature fusion to obtain the third fused feature; the third fused feature is passed to the decoder network in one path and to the MDFP module for downsampling to obtain the fifth downsampled feature; the fifth downsampled feature is concatenated with the P4 layer enhancement feature in the channel dimension to obtain the fourth concatenated feature;

[0040] The fourth concatenated feature is input into the wConv-C3k2 module for feature fusion to obtain the fourth fused feature; the fourth fused feature is finally input into the decoder network.

[0041] In one embodiment, the execution process of the wConv-C3k2 module is as follows:

[0042] The third input feature of the wConv-C3k2 module is split into channel dimensions using a 1×1 convolutional layer to obtain a first branch feature and a second branch feature; the third input feature is a first concatenation feature, a second concatenation feature, a third concatenation feature, or a fourth concatenation feature.

[0043] The features of the first branch are input into multiple cascaded weighted bottleneck modules for feature extraction, resulting in multi-level bottleneck features;

[0044] The multi-level bottleneck features and the second branch features are combined through channel splicing to obtain spliced ​​features;

[0045] The spliced ​​features are processed by channel fusion using a 1×1 convolutional layer to obtain a second output feature, which is either a first fused feature, a second fused feature, a third fused feature, or a fourth fused feature.

[0046] In one embodiment, the data processing procedure for the weighted bottleneck module is as follows:

[0047] The first branch features are processed by channel reduction and feature extraction based on the first weighted convolutional layer of the cascaded weighted bottleneck module to obtain the first weighted convolutional features.

[0048] Based on the second weighted convolutional layer of the weighted bottleneck module, channel recovery and feature extraction are performed on the first weighted convolutional features to obtain the second weighted convolutional features;

[0049] The first branch feature of the weighted bottleneck module is added element-wise to the second weighted convolution feature to obtain the third output feature of the weighted bottleneck module. The third output feature of the weighted bottleneck module is used as the input of the next level weighted bottleneck module so that the next level weighted bottleneck module performs the same operation and finally obtains multi-level bottleneck features.

[0050] In one embodiment, the data processing for the first weighted convolutional layer and the second weighted convolutional layer is the same:

[0051] A symmetric density vector is constructed based on preset density parameters to obtain a density coefficient vector;

[0052] The density weighting matrix is ​​obtained by performing an outer product operation on the density coefficient vector.

[0053] The learnable convolution kernel weight parameters are multiplied element-wise with the density weighting matrix to obtain the weighted convolution kernel weights.

[0054] The fourth input feature is obtained by performing a two-dimensional convolution operation on the fourth input feature using weighted convolution kernel weights; the fourth input feature is either the first branch feature or the first weighted convolution feature, and the fourth output feature is either the first weighted convolution feature or the second weighted convolution feature.

[0055] In one embodiment, the downsampling process of the MDFP module is as follows:

[0056] The fifth input feature is divided into three parallel branches, which are respectively processed by pixel rearrangement downsampling, convolution downsampling and max pooling downsampling; the fifth input feature is either the second fusion feature or the third fusion feature.

[0057] The first branch performs pixel splitting on the fifth input feature according to the spatial odd and even positions to obtain the first sub-feature, the second sub-feature, the third sub-feature, and the fourth sub-feature;

[0058] The first sub-feature, the second sub-feature, the third sub-feature, and the fourth sub-feature are concatenated to obtain the channel rearranged feature;

[0059] The channel rearrangement features are processed by channel fusion using a 1×1 convolutional layer to obtain fused features.

[0060] The fused features are normalized by a batch normalization layer to obtain pixel rearrangement downsampling features;

[0061] The second branch performs feature extraction processing on the fifth input feature through a 3×3 grouped convolutional layer with a stride of 1, to obtain the first depthwise convolutional feature.

[0062] The first depthwise convolutional features are downsampled by a 3×3 depthwise convolutional layer with a stride of 2 to obtain the convolutional downsampled features.

[0063] The convolutional downsampling features are activated using the GELUS activation function to obtain activated features;

[0064] The activation features are normalized by a batch normalization layer to obtain convolutional downsampling features;

[0065] The third branch downsamples the first depthwise convolutional features of the second branch through a max pooling layer with a stride of 2, resulting in pooled downsampled features.

[0066] The pooling downsampling features are normalized by a batch normalization layer to obtain the pooling downsampling features;

[0067] The pixel rearrangement downsampling features, convolution downsampling features and pooling downsampling features are processed by channel concatenation to obtain three-way concatenated features;

[0068] The three-way spliced ​​features are processed by channel fusion through a 1×1 convolutional layer to obtain the fifth output feature of the MDFP module. The fifth output feature is either the fourth downsampled feature or the fifth downsampled feature.

[0069] In one embodiment, the data processing procedure of the MDFP module is expressed by the following formula:

[0070]

[0071]

[0072]

[0073]

[0074] in, This represents the fifth input feature of the MDFP module; This refers to a 3×3 grouped convolutional layer with a stride of 1. This refers to a 3×3 depth convolutional layer with a stride of 2. This represents the GELUS activation function. This indicates the batch normalization layer. This represents the convolutional downsampling feature; This indicates the max pooling layer with a step size of 2. This represents the pooling downsampling feature; This refers to the first sub-feature, the second sub-feature, the third sub-feature, and the fourth sub-feature. This indicates the channel fusion processing operation. This represents the 1×1 convolutional layer. This represents the pixel rearrangement downsampling feature; This represents the fifth output feature.

[0075] In one embodiment, determining the prediction result based on the query interaction features and the multi-scale key-value features includes:

[0076] The query interaction features and the multi-scale key-value features are input into the deformable attention module, and adaptive spatial sampling is performed on the multi-scale feature map in the deformable attention module to obtain attention-enhanced query features.

[0077] The attention-enhanced query features are iteratively refined through multiple decoding layers to obtain the decoding output features of each layer;

[0078] The target category probability and normalized bounding box coordinates are calculated based on the decoding output features of each layer to obtain the prediction results for each layer.

[0079] This invention introduces the ODConv module to learn parallel attention across four dimensions of the convolution kernel space, dynamically modulates the convolution kernel, captures subtle features of the target, and effectively improves the ability to distinguish similar interference objects. It also constructs the wConv-C3k2 module, which dynamically adjusts the convolution kernel weights by introducing a spatial density function. This allows the model to prioritize the core features of the target, reduce the interference of background noise, and improve the quality of spatial feature extraction. Furthermore, it introduces the MDFP module, which extracts semantic, texture, and positional features through three complementary branches, preserving small target details without weighting, reducing the loss of key features during downsampling, significantly enhancing the small target detection capability, and slightly reducing the model's computational cost and parameter count. Attached Figure Description

[0080] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0081] in:

[0082] Figure 1 This is an application environment diagram of a target detection method based on the improved DEIM in complex field environments, as shown in one embodiment.

[0083] Figure 2 This is a flowchart of a target detection method based on an improved DEIM method in complex field environments, as described in one embodiment.

[0084] Figure 3 This is a schematic diagram of the ODConv module in one embodiment;

[0085] Figure 4 This is a schematic diagram illustrating the working principle of the wConv-C3k2 module in one embodiment;

[0086] Figure 5This is a schematic diagram of the MDFP module in one embodiment;

[0087] Figure 6 This is a structural block diagram of a computer device in one embodiment. Detailed Implementation

[0088] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0089] To address the technical problems in the background art, this application provides a target detection method based on an improved DEIM (Deep Detection Information Model) for complex outdoor environments.

[0090] Figure 1 This is an application environment diagram of a target detection method based on the improved DEIM in complex field environments, as shown in one embodiment. (Refer to...) Figure 1This improved target detection method based on DEIM (Deep Earth Detection and Analysis) in complex field environments is applied to an improved target detection system for complex field environments based on DEIM. The improved target detection system for complex field environments based on DEIM includes a terminal 110 and a server 120. The terminal 110 and server 120 are connected via a network. The terminal 110 can be a desktop terminal or a mobile terminal, and the mobile terminal can be at least one of a mobile phone, tablet computer, or laptop computer. The server 120 can be implemented using a standalone server or a server cluster consisting of multiple servers. Terminal 110 acquires images from the SARD dataset and performs data augmentation operations on the acquired images, including Mosaic stitching, random photometric distortion, random scaling, random IoU cropping, random horizontal flipping, and Mixup. The augmented images are then preprocessed to obtain the first input features. Server 120 receives the first input features and outputs P2, P3, and P4 layer features to the encoder network. The P2, P3, and P4 layer features are then processed by 1×1 convolutional layers through channel dimensionality reduction and projection to obtain the P2, P3, and P4 layer input features, respectively. These features are then input into the FPN path of the encoder network. The P4 layer input features are processed by the AIFI module in the FPN path. After layer feature enhancement, P4 layer enhanced features are obtained; and based on the P4 layer enhanced features, second, third, and fourth fusion features are obtained and input into the decoder network; the second, third, and fourth fusion features are received as multi-scale fusion features, and the multi-scale fusion features are used as key-value inputs for cross-attention to obtain multi-scale key-value features; the initial learnable target query vector is used as the initial query input, and noisy real-labeled auxiliary queries are introduced during the training phase to form a decoding query set; the decoding query set is input into the self-attention module of the decoding layer to perform information interaction processing between queries to obtain query interaction features; the prediction result is determined based on the query interaction features and the multi-scale key-value features; the prediction result is a natural human posture such as running, walking, standing, sitting, or lying down.

[0091] like Figure 2 As shown, in one embodiment, an improved target detection method based on DEIM in complex outdoor environments is provided. This method can be applied to both terminals and servers; this embodiment illustrates its application to a terminal. The improved target detection method based on DEIM in complex outdoor environments specifically includes the following steps:

[0092] S1: Obtain images from the SARD dataset, and apply data augmentation operations including Mosaic stitching, random photometric distortion, random scaling, random IoU cropping, random horizontal flipping, and Mixup to the obtained images. After preprocessing the data-augmented images, the first input features are obtained.

[0093] S2: The backbone network receives the first input feature and outputs the P2 layer feature, P3 layer feature and P4 layer feature to the encoder network;

[0094] S3: The encoder network performs channel dimensionality reduction and projection processing on the features of layer P2, layer P3, and layer P4 respectively through 1×1 convolutional layers to obtain the input features of layer P2, layer P3, and layer P4. These input features are then fed into the FPN path within the encoder network.

[0095] S4: The P4 layer input features are enhanced by the AIFI module in the FPN path to obtain the P4 layer enhanced features; and the second fusion feature, the third fusion feature, and the fourth fusion feature are obtained based on the P4 layer enhanced features and then input to the decoder network.

[0096] S5: Receive the second, third, and fourth fusion features as multi-scale fusion features, and use the multi-scale fusion features as key-value inputs for cross-attention to obtain multi-scale key-value features; use the initialized learnable target query vector as the initial query input, and introduce additional noisy real-labeled auxiliary queries during the training phase to form a decoded query set; input the decoded query set into the self-attention module of the decoding layer to perform information interaction processing between queries to obtain query interaction features; determine the prediction result based on the query interaction features and the multi-scale key-value features; the prediction result is a natural human posture such as running, walking, standing, sitting, or lying down.

[0097] In one embodiment, the backbone network includes: a StemBlock layer and four BasicStage layers, wherein the StemBlock layer is a feature extraction module; the backbone network receives the first input features and outputs P2 layer features, P3 layer features, and P4 layer features to the encoder network, including:

[0098] The first input features are initially downsampled by a 3×3 convolution with a stride of 2 in the feature extraction module to obtain the initial features;

[0099] The initial features are input into two branches. The initial features are subjected to max pooling in the first path to obtain salient features. The initial features are then subjected to two 2×2 convolutions in the second path to perform channel-dimensional feature transformation to obtain transformed features. The salient features and transformed features are then concatenated and fused to obtain fused features. The fused features are then downsampled and extracted again by a 3×3 convolution with a stride of 2 to obtain extracted features. The extracted features are then subjected to a 1×1 convolution for channel adjustment to obtain basic features.

[0100] The basic features are input into the first BasicStage layer. The multi-layer ODConv module in the HG-Block module of the first BasicStage layer extracts features layer by layer to obtain multiple first-layer features. The multiple first-layer features are then concatenated with the basic features to obtain multi-scale fusion features. The multi-scale fusion features are input into the SE compression-excited attention module in the first BasicStage layer for channel-dimensional feature recalibration, and the enhanced features are output to the second BasicStage layer.

[0101] In the second BasicStage layer, the input enhancement features are first downsampled by a 3×3 depthwise separable convolution with a stride of 2 to obtain the first downsampled features. The first downsampled features are then extracted layer by layer by the multi-layer ODConv module in the HG-Block module of the second BasicStage layer to obtain multiple second-layer features. The multiple second-layer features are then concatenated with the first downsampled features and aggregated by the SE compression excitation attention module to obtain the P2 layer features. The P2 layer features are then output to the third BasicStage layer of the encoder network.

[0102] In the third BasicStage layer, the P2 layer features are first downsampled by a 3×3 depthwise separable convolution with a stride of 2 to obtain the second downsampled features. The second downsampled features are then extracted layer by layer by the multi-layer ODConv module in the HG-Block module of the third BasicStage layer to obtain multiple third layer features. The multiple third layer features are then concatenated with the second downsampled features and aggregated by the SE compression and excitation attention module to obtain the P3 layer features. The P3 layer features are then output to the encoder network and the fourth BasicStage layer.

[0103] In the fourth BasicStage layer, the P3 layer features are first downsampled using a 3×3 depthwise separable convolution with a stride of 2 to obtain the third downsampled features. The third downsampled features are then extracted layer by layer by the multi-layer ODConv module within the HG-Block module in the fourth BasicStage layer to obtain multiple fourth layer features. These multiple fourth layer features are then concatenated with the third downsampled features and aggregated by the SE compression-excited attention module to obtain the P4 layer features. The P4 layer features are then output to the encoder network.

[0104] In one embodiment, such as Figure 3 As shown, the execution process of the multi-layer ODConv module is as follows:

[0105] The second input feature is subjected to global average pooling processing by an adaptive global average pooling layer to obtain a global pooling feature; the global pooling feature is then subjected to channel dimensionality reduction processing by a 1×1 convolutional layer to obtain a dimensionality-reduced feature; the dimensionality-reduced feature is then subjected to normalization processing and activation processing by a layer normalization layer and a ReLU activation layer in sequence to obtain a channel compression feature; the second input feature is a basic feature, a first downsampled feature, a second downsampled feature, or a third downsampled feature;

[0106] The channel attention weights are obtained by performing fully connected mapping and Sigmoid activation on the channel compressed features based on the channel attention branches.

[0107] The channel compression features are fully connected and Sigmoid activation processed according to the filter attention branch to obtain the filter attention weights.

[0108] Based on the spatial attention branch, the channel compression features are fully connected and Sigmoid activation is applied to obtain the spatial attention weights.

[0109] Based on the convolution kernel attention branch, the channel compression features are fully connected and Softmax normalized to obtain the convolution kernel attention weights.

[0110] The second input feature is multiplied element-wise according to the channel attention weights to obtain the channel-weighted feature;

[0111] Based on the spatial attention weights and the convolutional kernel attention weights, multiple sets of learnable convolutional kernel weight parameters are multiplied element-wise to obtain weighted multi-kernel weights.

[0112] The weighted multi-kernel weights are weighted and aggregated along the convolution kernel dimension to obtain aggregated convolution kernel weights.

[0113] The channel-weighted features are subjected to two-dimensional convolution processing based on the aggregated convolution kernel weights to obtain the initial convolution output features;

[0114] The initial convolutional output features are multiplied element-wise according to the filtering attention weights to obtain a first output feature, wherein the first output feature is a plurality of first-layer features, a plurality of second-layer features, a plurality of third-layer features, or a plurality of fourth-layer features.

[0115] The first output feature is achieved through the following expression:

[0116]

[0117] in, Indicates the second input feature, Indicates the first output feature. For multiple sets of learnable convolutional kernel weight parameters, , , as well as These represent the spatial attention weights, channel attention weights, filter attention weights, and convolutional kernel attention weights, respectively.

[0118] In one embodiment, obtaining the second fusion feature, the third fusion feature, and the fourth fusion feature based on the enhanced features of the P4 layer, and inputting them into the decoder network includes:

[0119] The enhanced features of layer P4 are transmitted in two paths. One path is transmitted to the PAN path, and the other path is upsampled and then concatenated with the input features of layer P3 along the channel dimension to obtain the first concatenated feature.

[0120] The first concatenated feature is input into the wConv-C3k2 module for feature fusion to obtain the first fused feature; the first fused feature is then subjected to a convolution operation to obtain the first convolutional feature;

[0121] The first convolutional feature is passed in two paths. One path is passed to the PAN path, and the other path is upsampled and then concatenated with the input feature of layer P2 along the channel dimension to obtain the second concatenated feature.

[0122] The second concatenated feature is input into the wConv-C3k2 module for feature fusion to obtain the second fused feature; the second fused feature is passed to the decoder network in one path and to the PAN path in the other path;

[0123] In the PAN path, the second fusion feature is downsampled by the MDFP module to obtain the fourth downsampled feature. The fourth downsampled feature is then concatenated with the P4 layer enhancement feature at the channel dimension to obtain the third concatenated feature.

[0124] The third concatenated feature is input into the wConv-C3k2 module for feature fusion to obtain the third fused feature; the third fused feature is passed to the decoder network in one path and to the MDFP module for downsampling to obtain the fifth downsampled feature; the fifth downsampled feature is concatenated with the P4 layer enhancement feature in the channel dimension to obtain the fourth concatenated feature;

[0125] The fourth concatenated feature is input into the wConv-C3k2 module for feature fusion to obtain the fourth fused feature; the fourth fused feature is finally input into the decoder network.

[0126] In one embodiment, such as Figure 4 As shown, the execution process of the wConv-C3k2 module is as follows:

[0127] The third input feature of the wConv-C3k2 module is split into channel dimensions using a 1×1 convolutional layer to obtain a first branch feature and a second branch feature; the third input feature is a first concatenation feature, a second concatenation feature, a third concatenation feature, or a fourth concatenation feature.

[0128] The features of the first branch are input into multiple cascaded weighted bottleneck modules for feature extraction, resulting in multi-level bottleneck features;

[0129] The multi-level bottleneck features and the second branch features are combined through channel splicing to obtain spliced ​​features;

[0130] The spliced ​​features are processed by channel fusion using a 1×1 convolutional layer to obtain a second output feature, which is either a first fused feature, a second fused feature, a third fused feature, or a fourth fused feature.

[0131] In one embodiment, the data processing procedure for the weighted bottleneck module is as follows:

[0132] The first branch features are processed by channel reduction and feature extraction based on the first weighted convolutional layer of the cascaded weighted bottleneck module to obtain the first weighted convolutional features.

[0133] Based on the second weighted convolutional layer of the weighted bottleneck module, channel recovery and feature extraction are performed on the first weighted convolutional features to obtain the second weighted convolutional features;

[0134] The first branch feature of the weighted bottleneck module is added element-wise to the second weighted convolution feature to obtain the third output feature of the weighted bottleneck module. The third output feature of the weighted bottleneck module is used as the input of the next level weighted bottleneck module so that the next level weighted bottleneck module performs the same operation and finally obtains multi-level bottleneck features.

[0135] In one embodiment, the data processing for the first weighted convolutional layer and the second weighted convolutional layer is the same:

[0136] A symmetric density vector is constructed based on preset density parameters to obtain a density coefficient vector;

[0137] The density weighting matrix is ​​obtained by performing an outer product operation on the density coefficient vector.

[0138] The learnable convolution kernel weight parameters are multiplied element-wise with the density weighting matrix to obtain the weighted convolution kernel weights.

[0139] A four-dimensional convolution operation is performed on the fourth input feature by weighting the convolution kernel weights to obtain the fourth output feature; the fourth input feature is either the first branch feature or the first weighted convolution feature, and the fourth output feature is either the first weighted convolution feature or the second weighted convolution feature;

[0140] The data processing procedure of the first weighted convolutional layer or the second weighted convolutional layer is expressed by the following formula:

[0141]

[0142]

[0143] in, For the density coefficient vector, The outer product of the density coefficient vector, This is the density weighting matrix obtained by performing an outer product operation on the density coefficient vector; The learnable convolutional kernel weight parameters, This is an element-wise multiplication operation. The weights are those of the weighted convolution kernel.

[0144] In one embodiment, such as Figure 5 As shown, the downsampling process of the MDFP module is as follows:

[0145] The fifth input feature is divided into three parallel branches, which are respectively processed by pixel rearrangement downsampling, convolution downsampling and max pooling downsampling; the fifth input feature is either the second fusion feature or the third fusion feature.

[0146] The first branch performs pixel splitting on the fifth input feature according to the spatial odd and even positions to obtain the first sub-feature, the second sub-feature, the third sub-feature, and the fourth sub-feature;

[0147] The first sub-feature, the second sub-feature, the third sub-feature, and the fourth sub-feature are concatenated to obtain the channel rearranged feature;

[0148] The channel rearrangement features are processed by channel fusion using a 1×1 convolutional layer to obtain fused features.

[0149] The fused features are normalized by a batch normalization layer to obtain pixel rearrangement downsampling features;

[0150] The second branch performs feature extraction processing on the fifth input feature through a 3×3 grouped convolutional layer with a stride of 1, to obtain the first depthwise convolutional feature.

[0151] The first depthwise convolutional features are downsampled by a 3×3 depthwise convolutional layer with a stride of 2 to obtain the convolutional downsampled features.

[0152] The convolutional downsampling features are activated using the GELUS activation function to obtain activated features;

[0153] The activation features are normalized by a batch normalization layer to obtain convolutional downsampling features;

[0154] The third branch downsamples the first depthwise convolutional features of the second branch through a max pooling layer with a stride of 2, resulting in pooled downsampled features.

[0155] The pooling downsampling features are normalized by a batch normalization layer to obtain the pooling downsampling features;

[0156] The pixel rearrangement downsampling features, convolution downsampling features and pooling downsampling features are processed by channel concatenation to obtain three-way concatenated features;

[0157] The three-way spliced ​​features are processed by channel fusion through a 1×1 convolutional layer to obtain the fifth output feature of the MDFP module. The fifth output feature is either the fourth downsampled feature or the fifth downsampled feature.

[0158] In one embodiment, the data processing procedure of the MDFP module is expressed by the following formula:

[0159]

[0160]

[0161]

[0162]

[0163] in, This represents the fifth input feature of the MDFP module; This refers to a 3×3 grouped convolutional layer with a stride of 1. This refers to a 3×3 depth convolutional layer with a stride of 2. This represents the GELUS activation function. This indicates the batch normalization layer. This represents the convolutional downsampling feature; This indicates the max pooling layer with a step size of 2. This represents the pooling downsampling feature; This refers to the first sub-feature, the second sub-feature, the third sub-feature, and the fourth sub-feature. This indicates the channel fusion processing operation. This represents the 1×1 convolutional layer. This represents the pixel rearrangement downsampling feature; Indicates the fifth output feature

[0164] In one embodiment, determining the prediction result based on the query interaction features and the multi-scale key-value features includes:

[0165] The query interaction features and the multi-scale key-value features are input into the deformable attention module, and adaptive spatial sampling is performed on the multi-scale feature map in the deformable attention module to obtain attention-enhanced query features.

[0166] The attention-enhanced query features are iteratively refined through multiple decoding layers to obtain the decoding output features of each layer;

[0167] The target category probability and normalized bounding box coordinates are calculated based on the decoding output features of each layer to obtain the prediction results for each layer.

[0168] Real-time acquisition of target image frames collected by the UAV in the field, and processing of the target image frames using the tested DEIM network to obtain the target detection results of the UAV. Experimental results on the SARD dataset show that, compared with the baseline model DEIM, the proposed method improves AP50:95 from 48.7% to 52.2%, AP75 from 51.1% to 56.0%, AP50 from 81.1% to 84.1%, and mAR from 78.8% to 84.2% with only a slight increase in computational cost and parameter count. These experimental results demonstrate the effectiveness of the proposed method.

[0169] Figure 6 An internal structural diagram of a computer device in one embodiment is shown. This computer device can specifically be a terminal or a server. Figure 6As shown, the computer device includes a processor, memory, and network interface connected via a system bus. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and may also store a computer program. When executed by the processor, this computer program enables the processor to implement an improved DEIM-based target detection method in complex field environments. The internal memory may also store a computer program, which, when executed by the processor, enables the processor to implement the improved DEIM-based target detection method in complex field environments. Those skilled in the art will understand that... Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0170] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. 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), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0171] 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.

[0172] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent 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 these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A target detection method based on an improved DEIM (Deep Detection Information Model) for complex field environments, characterized in that, The method includes: Images are obtained from the SARD dataset. The obtained images are subjected to data augmentation operations including Mosaic stitching, random photometric distortion, random scaling, random IoU cropping, random horizontal flipping, and Mixup. The first input features are obtained after preprocessing the data-augmented images. The backbone network receives the first input feature and outputs the P2 layer features, P3 layer features and P4 layer features to the encoder network; The encoder network performs channel dimensionality reduction and projection processing on the features of layer P2, layer P3, and layer P4 through 1×1 convolutional layers to obtain the input features of layer P2, layer P3, and layer P4, respectively. These input features are then fed into the FPN path within the encoder network. The P4 layer input features are enhanced by the AIFI module in the FPN path to obtain the P4 layer enhanced features; and based on the P4 layer enhanced features, the second fusion feature, the third fusion feature, and the fourth fusion feature are obtained and input into the decoder network. The second, third, and fourth fusion features are received as multi-scale fusion features, and these multi-scale fusion features are used as key-value inputs for cross-attention to obtain multi-scale key-value features. An initialized learnable target query vector is used as the initial query input, and during the training phase, noisy real-labeled auxiliary queries are introduced to form a decoded query set. The decoded query set is input to the self-attention module of the decoding layer for information interaction processing between queries to obtain query interaction features. A prediction result is determined based on the query interaction features and the multi-scale key-value features. The prediction result represents natural human postures such as running, walking, standing, sitting, and lying down.

2. The target detection method based on the improved DEIM in complex field environments according to claim 1, characterized in that, The backbone network includes a StemBlock layer and four BasicStage layers, wherein the StemBlock layer is a feature extraction module; the backbone network receives the first input features and outputs P2 layer features, P3 layer features, and P4 layer features to the encoder network, including: The first input features are initially downsampled by a 3×3 convolution with a stride of 2 in the feature extraction module to obtain the initial features. The initial features are input into two branches. The initial features are subjected to max pooling in the first path to obtain salient features. The initial features are then subjected to two 2×2 convolutions in the second path to perform channel-dimensional feature transformation to obtain transformed features. The salient features and transformed features are then concatenated and fused to obtain fused features. The fused features are then downsampled and extracted again by a 3×3 convolution with a stride of 2 to obtain extracted features. The extracted features are then subjected to a 1×1 convolution for channel adjustment to obtain basic features. The basic features are input into the first BasicStage layer. The multi-layer ODConv module in the HG-Block module of the first BasicStage layer extracts features layer by layer to obtain multiple first-layer features. The multiple first-layer features are then concatenated with the basic features to obtain multi-scale fusion features. The multi-scale fusion features are input into the SE compression-excited attention module in the first BasicStage layer for channel-dimensional feature recalibration, and the enhanced features are output to the second BasicStage layer. In the second BasicStage layer, the input enhancement features are first downsampled by a 3×3 depthwise separable convolution with a stride of 2 to obtain the first downsampled features. The first downsampled features are then extracted layer by layer by the multi-layer ODConv module in the HG-Block module of the second BasicStage layer to obtain multiple second-layer features. The multiple second-layer features are then concatenated with the first downsampled features and aggregated by the SE compression excitation attention module to obtain the P2 layer features. The P2 layer features are then output to the third BasicStage layer of the encoder network. In the third BasicStage layer, the P2 layer features are first downsampled by a 3×3 depthwise separable convolution with a stride of 2 to obtain the second downsampled features. The second downsampled features are then extracted layer by layer by the multi-layer ODConv module in the HG-Block module of the third BasicStage layer to obtain multiple third layer features. The multiple third layer features are then concatenated with the second downsampled features and aggregated by the SE compression and excitation attention module to obtain the P3 layer features. The P3 layer features are then output to the encoder network and the fourth BasicStage layer. In the fourth BasicStage layer, the P3 layer features are first downsampled using a 3×3 depthwise separable convolution with a stride of 2 to obtain the third downsampled features. The third downsampled features are then extracted layer by layer by the multi-layer ODConv module within the HG-Block module in the fourth BasicStage layer to obtain multiple fourth layer features. These multiple fourth layer features are then concatenated with the third downsampled features and aggregated by the SE compression-excited attention module to obtain the P4 layer features. The P4 layer features are then output to the encoder network.

3. The target detection method based on the improved DEIM in complex field environments according to claim 2, characterized in that, The execution process of the multi-layer ODConv module is as follows: The second input feature is subjected to global average pooling processing by an adaptive global average pooling layer to obtain global pooling features; the global pooling features are then subjected to channel dimensionality reduction processing by a 1×1 convolutional layer to obtain dimensionality-reduced features. The dimensionality-reduced features are sequentially normalized and activated using a layer normalization layer and a ReLU activation layer to obtain channel compression features; the second input feature is either a basic feature, a first downsampled feature, a second downsampled feature, or a third downsampled feature. The channel attention weights are obtained by performing fully connected mapping and Sigmoid activation on the channel compressed features based on the channel attention branches. The channel compression features are fully connected and Sigmoid activation processed according to the filter attention branch to obtain the filter attention weights. Based on the spatial attention branch, the channel compression features are fully connected and Sigmoid activation is applied to obtain the spatial attention weights. Based on the convolution kernel attention branch, the channel compression features are fully connected and Softmax normalized to obtain the convolution kernel attention weights. The second input feature is multiplied element-wise according to the channel attention weights to obtain the channel-weighted feature; Based on the spatial attention weights and the convolutional kernel attention weights, multiple sets of learnable convolutional kernel weight parameters are multiplied element-wise to obtain weighted multi-kernel weights. The weighted multi-kernel weights are weighted and aggregated along the convolution kernel dimension to obtain aggregated convolution kernel weights. The channel-weighted features are subjected to two-dimensional convolution processing based on the aggregated convolution kernel weights to obtain the initial convolution output features; The initial convolutional output features are multiplied element-wise according to the filtering attention weights to obtain a first output feature, wherein the first output feature is a plurality of first-layer features, a plurality of second-layer features, a plurality of third-layer features, or a plurality of fourth-layer features. The first output feature is achieved through the following expression: in, Indicates the second input feature, Indicates the first output feature. For multiple sets of learnable convolutional kernel weight parameters, , , as well as These represent the spatial attention weights, channel attention weights, filter attention weights, and convolutional kernel attention weights, respectively.

4. The target detection method based on the improved DEIM in complex field environments according to claim 1, characterized in that, The process of obtaining the second, third, and fourth fusion features based on the enhanced features of the P4 layer and inputting them into the decoder network includes: The enhanced features of layer P4 are transmitted in two paths. One path is transmitted to the PAN path, and the other path is upsampled and then concatenated with the input features of layer P3 along the channel dimension to obtain the first concatenated feature. The first concatenated feature is input into the wConv-C3k2 module for feature fusion to obtain the first fused feature; the first fused feature is then subjected to a convolution operation to obtain the first convolutional feature; The first convolutional feature is passed in two paths. One path is passed to the PAN path, and the other path is upsampled and then concatenated with the input feature of layer P2 along the channel dimension to obtain the second concatenated feature. The second concatenated feature is input into the wConv-C3k2 module for feature fusion to obtain the second fused feature; the second fused feature is passed to the decoder network in one path and to the PAN path in the other path; In the PAN path, the second fusion feature is downsampled by the MDFP module to obtain the fourth downsampled feature. The fourth downsampled feature is then concatenated with the P4 layer enhancement feature at the channel dimension to obtain the third concatenated feature. The third concatenated feature is input into the wConv-C3k2 module for feature fusion to obtain the third fused feature; the third fused feature is passed to the decoder network in one path and to the MDFP module for downsampling to obtain the fifth downsampled feature; the fifth downsampled feature is concatenated with the P4 layer enhancement feature in the channel dimension to obtain the fourth concatenated feature; The fourth concatenated feature is input into the wConv-C3k2 module for feature fusion to obtain the fourth fused feature; the fourth fused feature is finally input into the decoder network.

5. The target detection method based on the improved DEIM in complex field environments according to claim 4, characterized in that, The execution process of the wConv-C3k2 module is as follows: The third input feature of the wConv-C3k2 module is split into channel dimensions using a 1×1 convolutional layer to obtain a first branch feature and a second branch feature; the third input feature is a first concatenation feature, a second concatenation feature, a third concatenation feature, or a fourth concatenation feature. The features of the first branch are input into multiple cascaded weighted bottleneck modules for feature extraction, resulting in multi-level bottleneck features; The multi-level bottleneck features and the second branch features are combined through channel splicing to obtain spliced ​​features; The spliced ​​features are processed by channel fusion using a 1×1 convolutional layer to obtain a second output feature, which is either a first fused feature, a second fused feature, a third fused feature, or a fourth fused feature.

6. The target detection method based on the improved DEIM in complex field environments according to claim 5, characterized in that, The data processing procedure for the weighted bottleneck module is as follows: The first branch features are processed by channel reduction and feature extraction based on the first weighted convolutional layer of the cascaded weighted bottleneck module to obtain the first weighted convolutional features. Based on the second weighted convolutional layer of the weighted bottleneck module, channel recovery and feature extraction are performed on the first weighted convolutional features to obtain the second weighted convolutional features; The first branch feature of the weighted bottleneck module is added element-wise to the second weighted convolution feature to obtain the third output feature of the weighted bottleneck module. The third output feature of the weighted bottleneck module is used as the input of the next level weighted bottleneck module so that the next level weighted bottleneck module performs the same operation and finally obtains multi-level bottleneck features.

7. The target detection method based on the improved DEIM in complex field environments according to claim 6, characterized in that, The data processing for the first and second weighted convolutional layers is the same: A symmetric density vector is constructed based on preset density parameters to obtain a density coefficient vector; The density weighting matrix is ​​obtained by performing an outer product operation on the density coefficient vector. The learnable convolution kernel weight parameters are multiplied element-wise with the density weighting matrix to obtain the weighted convolution kernel weights. The fourth input feature is obtained by performing a two-dimensional convolution operation on the fourth input feature using weighted convolution kernel weights; the fourth input feature is either the first branch feature or the first weighted convolution feature, and the fourth output feature is either the first weighted convolution feature or the second weighted convolution feature.

8. The target detection method based on the improved DEIM in complex field environments according to claim 4, characterized in that, The downsampling process of the MDFP module is as follows: The fifth input feature is divided into three parallel branches, which are respectively processed by pixel rearrangement downsampling, convolution downsampling and max pooling downsampling; the fifth input feature is either the second fusion feature or the third fusion feature. The first branch performs pixel splitting on the fifth input feature according to the spatial odd and even positions to obtain the first sub-feature, the second sub-feature, the third sub-feature, and the fourth sub-feature; The first sub-feature, the second sub-feature, the third sub-feature, and the fourth sub-feature are concatenated to obtain the channel rearranged feature; The channel rearrangement features are processed by channel fusion using a 1×1 convolutional layer to obtain fused features. The fused features are normalized by a batch normalization layer to obtain pixel rearrangement downsampling features; The second branch performs feature extraction processing on the fifth input feature through a 3×3 grouped convolutional layer with a stride of 1, to obtain the first depthwise convolutional feature. The first depthwise convolutional features are downsampled by a 3×3 depthwise convolutional layer with a stride of 2 to obtain the convolutional downsampled features. The convolutional downsampling features are activated using the GELUS activation function to obtain activated features; The activation features are normalized by a batch normalization layer to obtain convolutional downsampling features; The third branch downsamples the first depth convolutional features of the second branch through a max pooling layer with a stride of 2, resulting in pooled downsampled features. The pooling downsampling features are normalized by a batch normalization layer to obtain the pooling downsampling features; The pixel rearrangement downsampling features, convolution downsampling features and pooling downsampling features are concatenated to obtain three-way concatenated features. The three-way spliced ​​features are processed by channel fusion through a 1×1 convolutional layer to obtain the fifth output feature of the MDFP module. The fifth output feature is either the fourth downsampled feature or the fifth downsampled feature.

9. The target detection method based on the improved DEIM in complex field environments according to claim 8, characterized in that, The data processing procedure of the MDFP module is expressed by the following formula: in, This represents the fifth input feature of the MDFP module; This refers to a 3×3 grouped convolutional layer with a stride of 1. This refers to a 3×3 depth convolutional layer with a stride of 2. This represents the GELUS activation function. This indicates the batch normalization layer. This represents the convolutional downsampling feature; This indicates the max pooling layer with a step size of 2. This represents the pooling downsampling feature; This refers to the first sub-feature, the second sub-feature, the third sub-feature, and the fourth sub-feature. This indicates the channel fusion processing operation. This represents the 1×1 convolutional layer. This represents the pixel rearrangement downsampling feature; This represents the fifth output feature.

10. The target detection method based on the improved DEIM in complex field environments according to claim 1, characterized in that, The determination of the prediction result based on the query interaction features and the multi-scale key-value features includes: The query interaction features and the multi-scale key-value features are input into the deformable attention module, and adaptive spatial sampling is performed on the multi-scale feature map in the deformable attention module to obtain attention-enhanced query features. The attention-enhanced query features are iteratively refined through multiple decoding layers to obtain the decoding output features of each layer; The target category probability and normalized bounding box coordinates are calculated based on the decoding output features of each layer to obtain the prediction results for each layer.

Images