Improved yolov11 network and space debris x-ray image detection method

Abstract

The application provides an improved YOLOv11 network and a space debris X-ray image detection method, and aims to solve the technical problem that although the existing YOLOv11 network performs excellently in a static scene, it performs poorly in processing dynamic data with time sequence changes, such as in the task of detecting and identifying space debris X-ray images, and cannot effectively capture the changes and movements of debris at different times, resulting in a decrease in detection accuracy. The improved YOLOv11 network provided by the application, aiming at the dynamic change characteristics of space debris, introduces MDTA modules and GDFN modules in the feature fusion network of the YOLOv11 network, optimizes the performance of the YOLOv11 network in processing dynamic scenes or time sequence data, and improves the local structure learning ability of debris detection. Meanwhile, the ASFF module is introduced, which dynamically weights and fuses features of different scales to solve the problem of partial overlap and occlusion of debris in space debris X-ray images, and improves the detection accuracy of the YOLOv11 network in a dense environment.

Description

TECHNICAL FIELD

[0001] The application relates to a space debris X-ray image detection method, in particular to an improved YOLOv11 network and a space debris X-ray image detection method. BACKGROUND

[0002] With the rapid development of space exploration and satellite technology, the number and types of space debris are increasing, which poses a great threat to the safety of space objects such as spacecraft and satellites. In order to effectively identify and track these space debris, many advanced detection technologies have emerged. The existing space debris detection methods mainly rely on traditional image processing methods and model-based detection methods. However, due to the dynamic changes of space debris in high altitude, complex background and high-density overlapping problems, traditional space debris detection methods face many challenges and are difficult to meet the needs of high-precision real-time detection.

[0003] In recent years, deep learning technology has made significant breakthroughs in the field of computer vision, especially in the field of object detection. The object detection algorithm based on YOLO (You Only Look Once) network has become a standard method for real-time object detection due to its excellent speed and accuracy. However, the existing object detection algorithm based on YOLO network still has certain limitations in dealing with complex dynamic scenes such as space debris. For example, Figure 1 As shown in the figure, the YOLOv11 network includes a backbone network 1, a feature fusion network 2 and a detection head 3, which significantly improves the efficiency and accuracy of feature extraction, reduces the parameter quantity and computational complexity, realizes higher detection accuracy, and greatly improves the training and inference speed, reduces the system delay. In addition, the object detection algorithm based on YOLOv11 network supports target detection, instance segmentation, image classification, pose estimation and rotating bounding box detection, etc. a variety of tasks, showing wide applicability. YOLOv11 network is also optimized for different deployment scenarios, suitable for resource-constrained edge devices, and can also run efficiently on high-performance computing platforms, providing strong support for real-time applications and large-scale deployment. Although YOLOv11 network performs well in static scenes, it performs poorly in handling dynamic data with time series changes, such as detecting and identifying space debris X-ray images, and cannot effectively capture the changes and movements of space debris at different times, resulting in decreased detection accuracy. SUMMARY

[0004] The purpose of the present application is to solve the technical problem that although the existing YOLOv11 network performs excellently in a static scene, it performs poorly in processing dynamic data with time sequence changes, such as in the task of detecting and identifying spatial debris X-ray images, and cannot effectively capture the changes and movements of debris at different times, resulting in a decrease in detection accuracy, and to provide an improved YOLOv11 network and a spatial debris X-ray image detection method.

[0005] In order to achieve the above-mentioned purpose, the technical scheme provided by the present application is as follows:

[0006] An improved YOLOv11 network, characterized in that it comprises a backbone network, an enhanced feature fusion network and a detection head unit connected in sequence.

[0007] The backbone network is used for extracting local features of different scales from the input image, generating feature maps of different scales, and performing multi-scale feature fusion on the feature maps of different scales to obtain fused features.

[0008] The enhanced feature fusion network is used for enhanced feature fusion of the fused features and the feature maps of different scales, and comprises m first feature extraction and fusion units connected in sequence and n second feature extraction and fusion units connected in sequence, m>1, n>1; the first feature extraction and fusion unit comprises an upsampling module Upsample, a concatenation module Concat and an MDTA-GDFN module connected in sequence; the second feature extraction and fusion unit comprises a first convolutional layer Conv, a concatenation module Concat and an MDTA-GDFN module connected in sequence; the MDTA-GDFN module comprises an MDTA module and a GDFN module connected in sequence; the input end of the MDTA module is connected to the output end of the corresponding concatenation module Concat, and is used to introduce a multi-domain time attention mechanism to effectively capture the time sequence changes of spatial debris; the GDFN module is used to introduce a gating mechanism and a depth separable convolution to focus on a specific area of the input image and learn the local image structure, and the output end of the GDFN module serves as the output end of the MDTA-GDFN module, outputting enhanced fusion features of different scales.

[0009] The second input end of the concatenation module Concat in the m first feature extraction and fusion units, the input end of the upsampling module Upsample in the first first feature extraction and fusion unit, and the second input end of the concatenation module Concat in the last second feature extraction and fusion unit are connected to the output end of the backbone network; the output end of the GDFN module in the mth first feature extraction and fusion unit is connected to the input end of the first convolutional layer Conv in the first second feature extraction and fusion unit.

[0010] The detection head unit includes n+1 storage modules, n+1 ASFF modules and a detection head output module; the input ends of the n+1 storage modules are connected as the input ends of the entire detection head unit, are connected to the output ends of the GDFN modules in the mth first feature extraction and fusion unit and the output ends of the GDFN modules in the n second feature extraction and fusion units respectively, and are used to receive and store different scale enhanced fusion features output by the corresponding GDFN modules respectively; the input ends of each ASFF module are connected to the output ends of the n+1 storage modules respectively, and are used to further fuse the n+1 different scale enhanced fusion features in different ASFF modules to obtain corresponding n+1 mixed enhanced features; the input ends of the detection head output module are connected to the output ends of the n+1 ASFF modules respectively, and are used to perform feature prediction according to the n+1 mixed enhanced features, and the output ends thereof are connected as the output ends of the detection head unit, and are used to output the predicted overall features.

[0011] Further, the enhanced feature fusion network adopts a PAN-FPN structure.

[0012] Further, the backbone network adopts a CSPDarknet53 structure.

[0013] Further, the backbone network includes five second convolution layers Conv, four C3K2 modules, an SPPF module and a C2PSA block.

[0014] The first second convolution layer Conv is used to receive an external spatial fragment X-ray image as the input of the backbone network; the first second convolution layer Conv, the second second convolution layer Conv, the first C3K2 module, the third second convolution layer Conv, the second C3K2 module, the fourth second convolution layer Conv, the third C3K2 module, the fifth second convolution layer Conv, the fourth C3K2 module, the SPPF module and the C2PSA block are sequentially arranged in the order of input and output, and are used to perform local feature extraction of different scales through five times of down-sampling, generate feature maps of different scales, and perform multi-scale feature fusion on the feature maps of different scales to obtain fusion features.

[0015] Further, m=2 and n=2.

[0016] The second input end of the concatenation module Concat in the first first feature extraction and fusion unit is connected to the output end of the third C3K2 module, and the second input end of the concatenation module Concat in the second first feature extraction and fusion unit is connected to the output end of the second C3K2 module; the input end of the up-sampling module Upsample in the first first feature extraction and fusion unit and the second input end of the concatenation module Concat in the second second feature extraction and fusion unit are connected to the output end of the C2PSA block respectively.

[0017] The output end of the GDFN module in the second first feature extraction fusion unit is connected to the input end of the first convolutional layer Conv in the first second feature extraction fusion unit.

[0018] In addition, the application also provides a space debris X-ray image detection method, which is characterized by comprising the following steps:

[0019] Step 1, constructing a training set based on space debris X-ray images;

[0020] Step 2, building the improved YOLOv11 network;

[0021] Step 3, inputting the space debris X-ray images in the training set into the improved YOLOv11 network for training to obtain the trained improved YOLOv11 network;

[0022] Step 4, collecting space debris X-ray images in space and inputting them into the trained improved YOLOv11 network, and completing the detection and recognition of the space debris X-ray images according to the output of the improved YOLOv11 network.

[0023] The application has the following beneficial effects compared with the prior art:

[0024] 1、The improved YOLOv11 network provided by the application, aiming at the dynamic change characteristics of space debris, introduces MDTA modules and GDFN modules in the feature fusion network of the YOLOv11 network, the MDTA module introduces a multi-domain time attention mechanism, effectively captures spatial and temporal information, optimizes the performance of the YOLOv11 network in processing dynamic scenes or time series data, the GDFN module uses a depth separable convolution to effectively encode spatial neighborhood information, improves the local structure learning ability of the debris detection, especially when processing debris in X-ray images, it can better extract local features, thereby realizing more efficient spatio-temporal feature learning. In addition, the improved YOLOv11 network of the application also introduces an ASFF module, which dynamically weights and fuses features of different scales to solve the problem of partial overlap and occlusion of debris in space debris X-ray images, and improves the detection accuracy of the YOLOv11 network in dense environments, especially in the case of serious mutual interference between debris, it can effectively improve the recall rate and detection accuracy.

[0025] 2、The space debris X-ray image detection method provided by the application can not only improve the detection accuracy of space debris, but also process dynamic time series data and complex overlap problems, adapt to complex space environments and mutual interference between debris, has high precision, high efficiency and good robustness, is suitable for real-time detection of space debris, and has a wide application prospect. Attached Figure Description

[0026] Figure 1 This is a diagram showing the overall architecture of an existing YOLOv11 network.

[0027] Figure 2 This is a diagram showing the overall architecture of the improved YOLOv11 network in an embodiment of the present invention.

[0028] Figure 3 The diagram shows the results of using the space debris X-ray image detection method of the present invention to perform detection on different space debris detection tasks, where (a) to (f) are schematic diagrams of the detection results of six different space debris detection tasks.

[0029] The specific labeling in the attached diagram is as follows:

[0030] 1-Backbone network; 2-Feature fusion network; 3-Detection head; 4-Enhanced feature fusion network; 5-Detection head unit. Detailed Implementation

[0031] To make the advantages and features of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0032] The YOLOv11 network is an advanced object detection and multi-task processing algorithm. This invention improves upon the existing YOLOv11 network and uses the improved YOLOv11 network as the core object detection framework to achieve higher accuracy in object detection. Existing YOLOv11 networks typically include a backbone network 1, a feature fusion network 2, and a detection head 3. This invention introduces an MDTA-GDFN module (Multi-Depth Convolution Head Transposed Attention-Gated Depth Convolution Feedforward Network) into the feature fusion network 2 of the existing YOLOv11 network. This enhances its ability to capture temporal information and encode spatial neighborhood information when processing dynamic scenes, such as spatial debris X-ray images, thereby improving the accuracy of spatial debris detection and the ability to process sequential images. Simultaneously, this invention adds an ASFF module (Adaptive Spatial Feature Fusion Module) to the detection head. This utilizes multi-domain temporal attention mechanisms, depthwise separable convolution, and adaptive spatial feature fusion technology, enabling the network to possess stronger feature learning capabilities and detection accuracy when processing complex backgrounds, overlapping debris, occlusion, and dynamically changing spatial debris X-ray images.

[0033] A method for detecting space debris in X-ray images based on an improved YOLOv11 network, specifically including the following steps:

[0034] Step 1: Construct a training set based on X-ray images of space debris.

[0035] Step 2, build the improved YOLOv11 network.

[0036] As shown in Figure 2 , the improved YOLOv11 network includes a backbone network 1, an enhanced feature fusion network 4, and a detection unit 5.

[0037] The backbone network 1 is the input of the entire improved YOLOv11 network, which is the same as the backbone network 1 of the existing YOLOv11 network, adopts the CSPDarknet53 structure, extracts feature maps of different scales of images through a large number of convolution operations, and then learns different levels of information of the image on the feature maps of different scales through a hierarchical structure. In this embodiment, the backbone network 1 includes five second convolution layers Conv, four C3K2 modules, an SPPF module, and a C2PSA block.

[0038] The five second convolution layers Conv are used to extract features of different scales in sequence; the C3K2 module is a specific implementation of the C3 (CrossStage Partial) structure, where K2 represents that two Bottleneck (residual units) are contained inside, and the C3 structure itself is derived from the design idea of CSPNet (Cross Stage Partial Network). By dividing the input features into two parts, one part is directly reserved, and the other part is spliced with the former after Bottleneck, the redundant calculation is effectively reduced, and the gradient flow capacity is improved. In C3K2, only 2 Bottleneck are used, compared with the conventional C3 structure, the calculation amount and model volume are further reduced under the premise of ensuring the feature extraction ability, which is suitable for deployment in resource-limited environments and helps to improve the model's representation ability for multi-scale targets. Figure 1

[0039] ​The SPPF module is a spatial pyramid feature fusion technology using multi-scale pooling, and its main purpose is to enhance the network's perception ability at different scales through multi-scale feature fusion. The C2PSA block is used to introduce a spatial attention mechanism to improve detection accuracy by enhancing the features of key regions. The second convolutional layer Conv is used as the input of the backbone network 1 to receive external spatial debris X-ray images. The first second convolutional layer Conv, the second second convolutional layer Conv, the first C3K2 module, the third second convolutional layer Conv, the second C3K2 module, the fourth second convolutional layer Conv, the third C3K2 module, the fifth second convolutional layer Conv, the fourth C3K2 module, the SPPF module, and the C2PSA block are sequentially arranged in the order of input and output, and are used to extract local features at different scales through five times of downsampling, generate feature maps at different scales, and perform multi-scale feature fusion on the feature maps at different scales to obtain fused features. The feature maps at different scales can capture spatial information at different scales and have good detection capability for spatial debris of different sizes. The C2PSA block is used as the output of the backbone network 1, and the obtained fused features are input into the subsequent enhanced feature fusion network 4. In addition, the second C3K2 module and the third C3K2 module in this embodiment also separately output a corresponding scale of features, which are used as input into the subsequent enhanced feature fusion network 4. In other embodiments of the present application, several C3K2 modules can be selected for separate output as needed, and the corresponding modules in the enhanced feature fusion network 4 are also adjusted accordingly.

[0040] The kernel size of the first second convolutional layer Conv k = 3, the stride s = 2, the padding size p = 1, wherein the stride s is used to control the downsampling rate and the output resolution, and the padding size p is used to balance the size of the input and output and preserve edge information.

[0041] The output size of the first second convolutional layer Conv is calculated according to the following formula:

[0042]

[0043] wherein, is the input size of the first second convolutional layer Conv.

[0044] Substitute the data:

[0045]

[0046] That is, the output size of the first second convolutional layer Conv is: 256x256x64.

[0047] The second second convolutional layer Conv (k=3, s=2, p=1) performs down-sampling, and the output size is 128x128x64. The C3K2 module is used for channel expansion, and the output size is 128x128x128. The third second convolutional layer Conv (k=3, s=2, p=1) is used for down-sampling again, and the output size is 64x64x128.

[0048] The SPPF module is subjected to multiple max-pooling (k=5, s=1), and the output size is 16x16x1024. The expression is as follows:

[0049]

[0050] wherein, is max-pooling, denotes a splicing operation, and F is an input tensor.

[0051] The enhanced feature fusion network 4 adopts a PAN-FPN structure (path augmented feature pyramid structure). The structure enhances the fusion of position details of a shallow layer and semantic information of a deep layer from bottom to top, so that the spatial fragment X-ray image detection can better capture the position details of a target.

[0052] The enhanced feature fusion network 4 is used for enhancing the fusion of the fusion features and the feature maps of different scales to obtain enhanced fusion features of different scales. In the embodiment, the enhanced feature fusion network 4 includes two first feature extraction and fusion units connected in sequence and two second feature extraction and fusion units connected in sequence. The first feature extraction and fusion unit includes an up-sampling module Upsample, a splicing module Concat and an MDTA-GDFN module connected in sequence. The second feature extraction and fusion unit includes a first convolutional layer Conv, a splicing module Concat and an MDTA-GDFN module connected in sequence. The up-sampling module Upsample, the splicing module Concat and the first convolutional layer Conv are the same as the existing YOLOv11 network. The up-sampling module Upsample is used for restoring the spatial resolution of the input features. The splicing module Concat is used for splicing different features along a specific dimension. The first convolutional layer Conv is used for extracting features of a corresponding scale, and the kernel size k =3, the step size s =2.

[0053] The input end of the upsampling module Upsample in the first first feature extraction fusion unit and the second input end of the concatenation module Concat in the second second feature extraction fusion unit are connected with the output end of the C2PSA block, for receiving the fusion features output by the backbone network 1 respectively; meanwhile, the second input end of the concatenation module Concat in the first first feature extraction fusion unit is connected with the output end of the third C3K2 module, and the second input end of the concatenation module Concat in the second first feature extraction fusion unit is connected with the output end of the second C3K2 module, for receiving the feature maps of corresponding scales output by different C3K2 modules in the backbone network 1 respectively. The output end of the MDTA-GDFN module in the second first feature extraction fusion unit is connected with the input end of the first convolution layer Conv in the first second feature extraction fusion unit.

[0054] One of the focuses of the present application is to replace the C3K2 module of the feature fusion network in the existing YOLOv11 network with the MDTA-GDFN module, and each MDTA-GDFN module includes a multi-depth convolution head transposed attention (MDTA, Multi-Deconvolution Head Transposed Attention) module and a gated deconvolution feed-forward network (GDFN, Gated Deconvolution Feed-forward Network) module connected in sequence. The input end of the MDTA module is connected with the output end of the corresponding concatenation module Concat, for introducing a multi-domain time attention mechanism, so as to effectively capture the timing changes of spatial fragments and improve the detection capability of the model for dynamic scenes and timing data; meanwhile, the MDTA module significantly reduces the computational complexity by applying the self-attention mechanism to the feature dimension instead of the spatial dimension, and improves the processing efficiency of high-resolution images. The GDFN module adopts a depth separable convolution, which can effectively encode the spatial adjacent pixel information and improve the learning ability of local image structures in the spatial fragment X-ray image, and the output end thereof serves as the output end of the MDTA-GDFN module, outputting enhanced fusion features of different scales.

[0055] The MDTA module effectively enhances the processing capability of the existing YOLOv11 network for timing changes, especially in dynamic scenes, by introducing a multi-domain time attention mechanism. The innovation of the MDTA module lies in applying the self-attention mechanism to the feature dimension, which can capture long-distance dependencies in the image, thereby improving the timing detection capability. The self-attention (SA) layer is usually used to undertake the computational overhead in the network, but the time and memory complexity of the key-query point product interaction in the SA layer grows quadratically with the spatial resolution of the input, i.e., for an image of WxH pixels (W and H are the width and height of the image respectively), the memory complexity is ​Therefore, it is infeasible to apply the SA layer to most image restoration tasks (usually involving high-resolution images). To solve this problem, the present application introduces the MDTA module, which has linear complexity for one reason that the MDTA module applies the SA layer to the channel dimension instead of the spatial dimension, i.e., the feature covariance across channels is computed to generate the attention map encoding the global context, and for another reason that a deep convolution is performed to emphasize the local context before the feature covariance is computed.

[0056] In a tensor of one layer normalization ( representing a three-dimensional real space with a size of HxWxC, H is the height of the image, W is the width, and C is the number of channels, representing all real numbers in the space), the MDTA module first generates query (Q), key (K) and value (V) projections, and aggregates the channel context across pixels through a 1x1 convolution, and then encodes the channel spatial context through a 3x3 deep convolution, to obtain Q projection, K projection and V projection respectively. Then, the Q projection and the K projection are reshaped so that their dot product interaction generates a transposed attention map with a size of instead of the regular attention map size . The overall processing process of the MDTA module is as follows:

[0057] ,

[0058] ,

[0059] wherein, is the output feature map, X is the input feature map, represents a 1x1 pointwise convolution operation, represents that the similarity of the query vector, the key vector and the value vector is calculated through the dot product, () represents that the normalization processing is performed, is a learnable scaling parameter, Q, K and V are the three basic components in the standard attention mechanism, representing query, key and value respectively.

[0060] The GDFN module combines deep separable convolution and gating mechanism. To transform features, a regular feedforward network (FN) usually performs independent and identical operations on each pixel location, which uses two 1x1 convolution layers, one 1x1 convolution layer is used to expand the feature channels, usually the expansion ratio of the feature channels is 4, that is, the number of processed feature channels will be 4 times the number of input feature channels, and the other 1x1 convolution layer is used to restore the feature channel number to the original input dimension. In order to apply nonlinear activation in the hidden layer, the application introduces gating mechanism and deep separable convolution through GDFN module, realizes two basic modifications of feedforward network, and improves the method of representation learning. Among them, the gating mechanism is defined as the element-wise product of two linear transformation layers in parallel paths, one of which uses GELU activation function for nonlinear activation, which can selectively focus on specific areas of the input image and improve the processing ability of image details. And the deep separable convolution is used to encode the information of spatially adjacent pixel positions, which helps to learn local image structure, realize effective recovery, significantly reduce the amount of calculation, and at the same time maintain the effectiveness of convolution operation.

[0061] Deep separable convolution:

[0062]

[0063] Gating mechanism:

[0064]

[0065] Output merging:

[0066]

[0067] wherein, is the generated convolution kernel weight of dynamic convolution, F is the input feature map, represents global average pooling on the input feature map, represents that the feature after global average pooling processing is taken as input, and the convolution kernel weight is learned through multi-layer network; represents the gating signal, which is used to determine whether the feature is activated, represents the gating signal after convolution operation, is the merged output feature map.

[0068] Given the input tensor , the overall processing process of the GDFN module is as follows:

[0069]

[0070]

[0071] wherein, is a weight matrix used to adjust the output of the gating mechanism, is an activation function used to introduce nonlinearity, , is a weight matrix used to adjust the input features and the features after layer normalization respectively, is a layer normalization operation, which is usually used to stabilize the training of the network, reduce the internal covariate shift, and improve the convergence of the model, represents a gating operation, is an intermediate result related to the operation of represents the output after transformation or activation, represents matrix multiplication.

[0072] The present application significantly improves the processing efficiency of the YOLOv11 network by combining the MDTA module and the GDFN module, and at the same time improves the comprehensive learning ability of the YOLOv11 network for local features and global context information.

[0073] The detection head unit 4 is located at the head of the target detection head, which adopts a decoupled structure, one branch of which is used to predict the target class, and the other branch is used to predict the target position. Through this decoupled design, the classification accuracy and positioning accuracy are optimized respectively.

[0074] In the embodiment, the detection head unit 4 includes three storage modules, three ASFF modules and a detection head output module. The input ends of the three storage modules are respectively used as the input ends of the entire detection head unit, and are connected to the output ends of the GDFN modules in the second first feature extraction and fusion unit and the two second feature extraction and fusion units in the enhanced feature fusion network 4, respectively, for receiving the different scale enhanced fusion features Level-1, Level-2 and Level-3 output by the corresponding GDFN modules and storing them; the three ASFF modules are respectively denoted as ASFF-1, ASFF-2 and ASFF-3, the input ends of ASFF-1, ASFF-2 and ASFF-3 are connected to the output ends of the three storage modules, respectively, for further fusing Level-1, Level-2 and Level-3 to obtain three mixed enhanced features; the input end of the detection head output module YOLO Detect is connected to the output ends of ASFF-1, ASFF-2 and ASFF-3, for performing feature prediction according to the mixed enhanced features output by ASFF-1, ASFF-2 and ASFF-3, and the output end thereof is used as the output end of the detection head unit, for outputting the predicted overall features.

[0075] The debris parts in the space debris X-ray image are often overlapped or blocked, although the multi-scale features can effectively handle objects of different sizes, but due to the inconsistency of features between scales, it is easy to cause gradient conflict and information loss in the training process, so that the traditional target detection method is difficult to effectively identify the target in these complex situations. Especially when using a single detector, when the detector maps the features of large objects and small objects to different levels, it is easy to introduce inconsistent features, thereby affecting the final detection performance. In order to improve the detection accuracy of YOLOv11 network in these environments, the application introduces an ASFF module, which dynamically fuses features from different scales through dynamic weighting, effectively suppresses feature conflict, improves the scale invariance of features, and thus significantly improves the performance of space debris detection.

[0076] The ASFF module first aligns the features of different scales through upsampling or downsampling, and then generates dynamic fusion weights to dynamically adjust the contribution of features according to the task and position, so that multi-scale features can better express target information. This dynamic weighted fusion method enhances the detection accuracy of YOLOv11 network for space debris in dense environments. The ASFF module can effectively fuse features from different scales by adaptively learning the spatial importance fusion weights of each layer of features, thereby improving the target detection accuracy, especially in processing multi-scale targets and complex backgrounds, which has a significant performance improvement.

[0077] The ASFF module realizes dynamic weighted fusion of multi-scale features through the following two key steps:

[0078] Feature scaling: In order to make features of different scales be able to be fused in the same scale, first, the features of different scales are scaled to realize alignment. For each layer of features , the ASFF method adjusts the features from other scales to the same resolution of the target scale . This process includes:

[0079] Upsampling: Adjust the channel number through 1x1 convolution, and then use interpolation method to increase the resolution of the feature.

[0080] Downsampling: Use convolution and step size of 2 to adjust the resolution and channel number.

[0081] Adaptive fusion: In each spatial position , the ASFF module fuses features from different scales by adaptively calculating the fusion weight. The specific formula is:

[0082]

[0083] wherein, is the fused feature, , and are the fusion weights learned adaptively on different scale features, representing the fusion importance of each scale feature at spatial position . , , are the feature representations of three different levels respectively.

[0084] Weight normalization: The ASFF module uses the Softmax operation to normalize the fusion weights, ensuring that all scale features are effectively balanced during feature fusion, avoiding the dominance of certain scale feature information, which may lead to information loss.

[0085] Weight calculation: The fusion weights , , are calculated by the following formula respectively:

[0086]

[0087]

[0088]

[0089] wherein, , and are the control parameters learned by the network, e （） is the exponential function.

[0090] The fusion weights , , are normalized by the Softmax function to ensure that the sum of all fusion weights is 1, that is:

[0091]

[0092] This calculation method can flexibly adjust the contribution of different scale features at a specific spatial position, effectively solving the feature conflict problem.

[0093] The introduction of the ASFF module solves the problem of detecting spatial debris in dense environments and improves the performance of the YOLOv11 network in complex scenes.

[0094] Step 3: Input the spatial debris X-ray images in the training set into the improved YOLOv11 network for training, and obtain the trained improved YOLOv11 network.

[0095] Step 4, a space debris X-ray image in the space is collected and input into the improved YOLOv11 network, and according to the output of the improved YOLOv11 network, the position and category of the space debris in the space debris X-ray image are determined, and the space debris X-ray image detection and recognition is completed.

[0096] The present application improves the YOLOv11 network and makes significant technical breakthroughs in many aspects:

[0097] 1. Improved space debris detection accuracy and recall rate: Compared with traditional target detection methods, the improved YOLOv11 network significantly improves the detection performance of space debris in complex environments.

[0098] 2. Enhanced processing capability for time series data: The MDTA module introduces a multi-domain time attention mechanism, enabling the model to effectively capture changes in debris at different times and enhance the model's target recognition capability in dynamic scenarios. This is crucial for processing time series data in space debris X-ray images, improving detection accuracy and reducing the rate of missed detections.

[0099] 3. Optimized multi-scale feature fusion: The ASFF module dynamically weights and fuses features from different scales, further improving the efficiency of multi-scale image feature fusion. This module can effectively handle the mutual interference between different types of space debris, especially in dense environments, improving detection accuracy and recall rate.

[0100] 4. Improved computational efficiency: The MDTA module applies self-attention mechanisms to feature dimensions and uses depth separable convolution, significantly reducing computational complexity, enabling the detection method to not only have high detection accuracy but also efficiently process high-resolution images.

[0101] In summary, the space debris X-ray image detection method of the present application has excellent performance in space debris detection, solving multiple key problems in the prior art and having wide application prospects and practical value.

[0102] The effects of the present application are verified through specific experiments.

[0103] During the experiment, the data set used was simulated space debris X-ray images, which included 14 categories, each category containing 500 514*514 grayscale images. Each category contained randomly positioned, sized, and oriented simulated images. There was also a group of multi-category superimposed images, which were 14 category images combined by random superposition, totaling 7467 images.

[0104] To improve the generalization ability of the improved YOLOv11 network, the present application adopts various data augmentation techniques such as rotation, scaling, cropping and mirroring during training. The data set is divided into 70% training set and 30% validation set to ensure that the improved YOLOv11 network can be trained and verified on different types of images. The training adopts the SGD optimization algorithm, and the learning rate adopts the dynamic adjustment strategy to accelerate the convergence and avoid overfitting. Every 64 Batch is verified once, and the performance of the improved YOLOv11 network is monitored through the validation set to ensure that the improved YOLOv11 network reaches the optimal state during training. The training period is set to 300, and the cross-validation technique is used to further optimize the parameters of the improved YOLOv11 network.

[0105] The accuracy (Accuracy), recall (Recall), mean average precision (mAP) and other common performance indicators are used to evaluate the effect of the improved YOLOv11 network. The mAP50 evaluation indicator is used to comprehensively understand the performance of the improved YOLOv11 network, especially the detection accuracy and recall rate in the detection of multiple classes of spatial fragments.

[0106]

[0107]

[0108]

[0109] AP: For each class, calculate the area under the precision-recall curve (AUC), and N is the number of total classes.

[0110] mAP: The average of all class AP values.

[0111] mAP50: A variant of mAP that only considers predictions with an IoU (Intersection over Union) greater than 0.5. This means that only when the overlap between the predicted box and the true box exceeds 50% will it be considered as a correct prediction.

[0112] As Figure 3 shown, the results of different spatial fragment detection tasks using the spatial fragment X-ray image detection method based on the improved YOLOv11 network of the present application are shown in the results of the six different spatial fragment detection tasks, as shown in (a)~(f), it can be seen that the detection method of the present application can accurately detect spatial fragment X-ray images in different spatial fragment detection tasks.

[0113] Compared with the existing YOLOv9 network, the improved YOLOv11 network of the application has significant improvement in detection accuracy, recall rate and other indicators. The detection results of the existing YOLOv9 network, the existing YOLOv11 network and the improved YOLOv11 network of the application are shown in Table 1:

[0114] Table 1: Space debris detection results of three different detection networks

[0115]

[0116] As can be seen, compared with the existing YOLOv9 network and the existing YOLOv11 network, the improved YOLOv11 network of the application has a detection accuracy and recall rate of more than 99%, excellent performance, the overall detection accuracy reaches 99.6%, the recall rate is 99.5%, the F1 score and the mAP50 value are also effectively improved, especially in the case of serious overlap and occlusion of debris, the improved YOLOv11 network can still maintain high detection accuracy.

[0117] The improved YOLOv11 network of the application can maintain high performance in complex X-ray space debris detection, especially in dynamic scenes and debris overlap, and has better robustness and stability.

[0118] In summary, the application innovatively improves the existing YOLOv11 network, combines the MDTA module, the GDFN module and the ASFF module, solves a number of problems in space debris X-ray image detection, and demonstrates significant advantages in high-precision detection, complex environment adaptation and time series data processing.

[0119] The above is only used to illustrate the technical solutions of the application, not to limit it. For ordinary professional technicians in the art, the specific technical solutions described in the above embodiments can be modified, or some technical features can be replaced, and these modifications or replacements do not make the corresponding technical solutions deviate from the scope of the technical solutions protected by the application.

Claims

1. A method for improving a YOLOv11 network, characterized in that: comprising a backbone network (1), an enhanced feature fusion network (4) and a detection head unit (5) connected in sequence; the backbone network (1) is used for local feature extraction of different scales on an input image, generating feature maps of different scales, and performing multi-scale feature fusion on the feature maps of different scales to obtain fused features; the enhanced feature fusion network (4) is used for enhanced feature fusion of the fused features and the feature maps of different scales, and comprises m first feature extraction and fusion units connected in sequence and n second feature extraction and fusion units connected in sequence, m≥1, n≥1; the first feature extraction and fusion unit comprises an upsampling module Upsample, a concatenation module Concat and an MDTA-GDFN module connected in sequence; the second feature extraction and fusion unit comprises a first convolutional layer Conv, a concatenation module Concat and an MDTA-GDFN module connected in sequence; the MDTA-GDFN module comprises an MDTA module and a GDFN module connected in sequence; the input end of the MDTA module is connected to the output end of the corresponding concatenation module Concat, used to introduce a multi-domain time attention mechanism to effectively capture the timing changes of spatial fragments; the GDFN module is used to introduce a gating mechanism and a depth separable convolution to focus on specific regions of the input image and learn local image structures, and the output end of the GDFN module is the output end of the MDTA-GDFN module, outputting enhanced fusion features of different scales; the second input end of the concatenation module Concat in the m first feature extraction and fusion units, the input end of the upsampling module Upsample in the first first feature extraction and fusion unit, and the second input end of the concatenation module Concat in the last second feature extraction and fusion unit are respectively connected to the output end of the backbone network (1); the output end of the GDFN module in the mth first feature extraction and fusion unit is connected to the input end of the first convolutional layer Conv in the first second feature extraction and fusion unit; the detection head unit (5) comprises n+1 storage modules, n+1 ASFF modules and a detection head output module; the input ends of the n+1 storage modules are connected to the input ends of the entire detection head unit, and are respectively connected to the output ends of the GDFN modules in the mth first feature extraction and fusion unit and the n second feature extraction and fusion units, used to respectively receive and store the enhanced fusion features of different scales output by the corresponding GDFN modules; the input ends of each ASFF module are respectively connected to the output ends of the n+1 storage modules, used to further fuse the n+1 enhanced fusion features of different scales in different ASFF modules to obtain corresponding n+1 mixed enhanced features; the input ends of the detection head output module are respectively connected to the output ends of the n+1 ASFF modules, used to perform feature prediction according to the n+1 mixed enhanced features, and the output end of the detection head output module is the output end of the detection head unit (5), used to output the predicted overall features. ​ ​ ​ 2. The method of claim 1, wherein the improved YOLOv11 network is characterized in that: the enhanced feature fusion network (4) adopts a PAN-FPN structure.

3. The method of claim 2, wherein the improved YOLOv11 network is characterized in that: the backbone network (1) adopts a CSPDarknet53 structure.

4. The method of claim 3, wherein the improved YOLOv11 network is characterized in that: the backbone network (1) includes five second convolution layers Conv, four C3K2 modules, an SPPF module, and a C2PSA block; a first second convolution layer Conv serves as an input of the backbone network (1) and is configured to receive an external spatial fragment X-ray image; a first second convolution layer Conv, a second second convolution layer Conv, a first C3K2 module, a third second convolution layer Conv, a second C3K2 module, a fourth second convolution layer Conv, a third C3K2 module, a fifth second convolution layer Conv, a fourth C3K2 module, an SPPF module, and a C2PSA block are sequentially arranged in an input-output order and are configured to perform local feature extraction at different scales through five times of down-sampling to generate feature maps at different scales, and perform multi-scale feature fusion on the feature maps at different scales to obtain fused features.

5. The method of claim 4, wherein the improved YOLOv11 network is characterized in that: m = 2 and n = 2; a second input end of a concatenation module Concat in a first first feature extraction and fusion unit is connected to an output end of the third C3K2 module, and a second input end of a concatenation module Concat in a second first feature extraction and fusion unit is connected to an output end of the second C3K2 module; an input end of an up-sampling module Upsample in the first first feature extraction and fusion unit and a second input end of a concatenation module Concat in the second second feature extraction and fusion unit are respectively connected to an output end of the C2PSA block; an output end of a GDFN module in the second first feature extraction and fusion unit is connected to an input end of a first convolution layer Conv in the first second feature extraction and fusion unit.

6. A method of detecting X-ray images of space debris, characterized in that, The method includes the following steps: Step 1, constructing a training set based on spatial fragment X-ray images; Step 2, building an improved YOLOv11 network according to the method of any one of claims 1-5; Step 3, inputting the spatial fragment X-ray images in the training set into the improved YOLOv11 network for training to obtain a trained improved YOLOv11 network; Step 4, collecting spatial fragment X-ray images in space and inputting them into the trained improved YOLOv11 network, and completing detection and recognition of the spatial fragment X-ray images according to an output of the improved YOLOv11 network.

Images