一种适用于低光空间环境的轻量级目标检测方法

Abstract

本发明属于航天科学技术领域，具体涉及一种适用于低光空间环境的轻量级目标检测方法，包括：步骤一：基于MobileNetV4构建轻量特征提取骨干网络；步骤二：结合TopkSPPF多分支池化实现多尺度特征精细融合，同时引入C2PSA注意力模块提升目标聚焦能力；步骤三：在特征融合阶段嵌入EdgeFca模块，融合小波边缘提取与频域加权机制，增强低光照下特征表达；步骤四：设计DNiN解耦检测头并完成模型训练优化。本发明解决现有技术在低光照空间环境下存在的目标检测精度低、边缘模糊、抗噪能力差以及不适用于星载边缘平台等问题，本发明具备频域建模能力与边缘增强机制，能够在资源受限的条件下实现对航天器、空间碎片等非合作目标的实时识别与感知。

Description

Technical Field

[0001] This invention belongs to the field of aerospace science and technology, specifically relating to a lightweight target detection method suitable for low-light space environments. Background Technology

[0002] With the continuous expansion of human space activities, the near-Earth orbit environment is becoming increasingly crowded, with a sharp increase in the number of active spacecraft and space debris, significantly raising the risk of collisions. Therefore, space target detection has become one of the key technologies for ensuring the safety of space missions. Currently, target detection based on monocular vision is widely used in space missions due to its advantages such as compact structure, low power consumption, and flexible deployment, and is particularly suitable for resource-constrained on-orbit platforms. However, most traditional visual detection methods rely on good lighting conditions, and their detection performance drops significantly in lighting-constrained scenarios such as extremely low light levels or backlighting in space.

[0003] Existing low-light target detection methods can be mainly divided into two categories: low-light image illumination correction methods and robust feature extraction methods. While the former can improve image visual quality, it easily introduces false edges and noise, affecting subsequent detection accuracy. The latter uses illumination-invariant features and synthetic data training methods, which can improve robustness, but rely on paired training samples or have high computational costs, making it difficult to meet the needs of real-time spaceborne deployment. Furthermore, commonly used ground-based low-light feature extractors perform poorly in complex space backgrounds because they struggle to simultaneously suppress high-frequency noise and preserve key edge structures, and they do not consider the importance of frequency domain characteristics in low-light images.

[0004] Therefore, there is an urgent need to design a target detection method that is suitable for low-light spatial scenes and has edge enhancement capabilities, frequency domain perception capabilities, and lightweight model characteristics, so as to improve the detection accuracy, boundary localization capabilities, and on-orbit deployment feasibility under extreme low light conditions. To this end, the present invention provides a lightweight target detection method suitable for low-light spatial environments. Summary of the Invention

[0005] The purpose of this invention is to provide a lightweight target detection method suitable for low-light space environments, which can solve the problems of low target detection accuracy, blurred edges, poor noise resistance, and unsuitability for spaceborne edge platforms in existing technologies under low-light space environments; it has frequency domain modeling capabilities and edge enhancement mechanisms, and can realize real-time identification and perception of non-cooperative targets such as spacecraft and space debris under resource-constrained conditions.

[0006] The specific technical solution adopted by this invention is as follows:

[0007] A lightweight target detection method suitable for low-light spatial environments includes the following steps:

[0008] Step 1: Construct a lightweight feature extraction backbone network based on MobileNetV4; Specifically: Design a lightweight feature extraction backbone network that balances accuracy and computational efficiency. A hierarchical network is built based on the MobileNetV4 architecture. In the initial layers of the detection framework, a computationally intensive convolutional network is used to achieve high computational efficiency with fewer channels while preserving detailed features. In the downsampling, small-size feature extraction stage, a memory-intensive depthwise separable convolution is used to balance the increased number of channels and reduce the computational complexity of convolution.

[0009] Step Two: Combine TopkSPPF multi-branch pooling to achieve fine-grained fusion of multi-scale features, while introducing a C2PSA attention module to enhance target focusing capabilities. Specifically: Add the multi-peak spatial pyramid pooling module TopkSPPF and the cross-layer spatial attention module C2PSA to achieve fine-grained fusion of multi-scale features and efficient weighting of cross-layer spatial features. The TopkSPPF module replaces the traditional SPPF module, aggregating multi-peak feature values ​​and refining feature representation through a multi-branch structure of double max pooling and average pooling. A multi-head attention mechanism is designed, utilizing grouped convolutions to calculate the Query, Key, and Value matrices and positional encoding. This attention mechanism enhances the focusing capability on the target region and suppresses background noise interference.

[0010] Step 3: Embedding the EdgeFca module in the feature fusion stage, integrating wavelet edge extraction and frequency domain weighting mechanisms to enhance feature representation under low illumination: Deploying multi-scale feature aggregation and edge enhancement modules to optimize low-light feature representation. The overall feature fusion framework adopts a lightweight inverted bottleneck module, embedding the EdgeFca module based on edge enhancement in the frequency domain to enhance feature representation using frequency domain information. Specifically, the EdgeFca module first extracts edge features through Haar wavelet transform, then enhances them through edge convolution and frequency domain attention weighting, and finally outputs robust features through dynamic residual fusion, achieving edge enhancement and noise suppression.

[0011] Step 4: Design the DNiN decoupled detection head and complete model training and optimization; specifically: design the decoupled dual-branch detection head DNiN, which outputs the target category and location label boxes. A lightweight design is achieved using depthwise separable convolution, and pixel-level sliding window detection is implemented using pointwise convolution. The detection head adopts a dual-branch parallel structure, using different loss functions for classification and label box regression respectively. Finally, it is trained on a constructed low-light spatial dataset to obtain a lightweight target detection framework suitable for low-light spatial environments.

[0012] This invention addresses the need for spaceborne target detection in low-light space environments by proposing a lightweight detection method that can efficiently extract the location and category information of low-light space targets. The beneficial effects of this invention are:

[0013] 1. Compared with traditional low-light detection technology, this invention proposes TopkSPPF and EdgeFca modules to collaboratively optimize feature representation, achieving a detection accuracy of 97.1% mAP50 and 77.3% mAP50-95 in low-light spatial scenes, and improving the recall rate for target recognition and classification to over 92%, significantly enhancing detection robustness.

[0014] 2. Compared with conventional complex detection models, this invention uses a lightweight architecture to reduce the computational load to as low as 2.9 BFLOPs, with a parameter scale of 1.15M, and the CPU inference speed of a single image can reach more than 35 FPS. It perfectly adapts to the resource constraints of the spaceborne edge platform and meets the real-time detection requirements.

[0015] 3. Compared with ground-based low-light detection methods, this invention is trained on a dedicated low-light spatial dataset and demonstrates excellent generalization ability on a self-built low-light dataset and the public datasets SPEED and SPEED+. Attached Figure Description

[0016] Figure 1 This is a flowchart of a lightweight target detection framework method suitable for low-light spatial environments according to the present invention;

[0017] Figure 2 This is a structural diagram of the lightweight feature extraction backbone network of the present invention;

[0018] Figure 3 This is a structural diagram of the TopkSPPF multi-peak spatial pyramid pooling module of the present invention;

[0019] Figure 4 This is a structural diagram of the cross-layer spatial attention module C2PSA of the present invention;

[0020] Figure 5 This is a structural diagram of the feature fusion neck network of the present invention;

[0021] Figure 6 This is a schematic diagram of the EdgeFca frequency domain attention module based on edge enhancement of the present invention;

[0022] Figure 7 This is a structural diagram of the decoupled dual-branch detection head DNiN of the present invention. Detailed Implementation

[0023] To make the objectives and advantages of this invention clearer, the invention will be specifically described below with reference to embodiments. It should be understood that the following text is merely used to describe one or more specific embodiments of the invention and does not strictly limit the scope of protection specifically claimed by the invention.

[0024] like Figures 1-7 As shown, a lightweight target detection framework method suitable for low-light spatial environments includes the following steps:

[0025] Step 1: Construct a lightweight feature extraction backbone network based on MobileNetV4, as detailed below:

[0026] like Figure 2 As shown, the backbone network adopts the MobileNetV4 architecture. The front-end uses three sets of alternating joint convolutional structures with kernel sizes of 3 and 1 to downsample the input image, progressively extracting semantic features and generating a feature map P3 with a spatial resolution one-eighth that of the original image. Each convolutional operation is followed by batch normalization and the SiLU activation function to enhance nonlinear expressiveness and accelerate network convergence. The SiLU activation function is continuously differentiable over the entire real number domain, and its expression is as follows:

[0027]

[0028] In the later stages of the network, Inverted Bottleneck and Extra Depthwise modules are used, replacing traditional convolutions with depthwise separable convolutions to effectively reduce the number of parameters and computational complexity. To avoid information loss due to truncating negative values ​​in low-dimensional space, no activation function is used in the convolutional layers at the end of all Inverted Bottleneck modules. In the residual connection strategy, identity mapping connections are preserved when the convolution stride = 1 to promote gradient flow; when the stride = 2, residual connections are canceled to complete spatial downsampling. Through the above structural combination, multi-scale feature maps P3, P4, and P5 are generated sequentially, corresponding to feature layers scaled at 1 / 8, 1 / 16, and 1 / 32 ratios, respectively.

[0029] Step 2: Combine TopkSPPF multi-branch pooling to achieve fine fusion of multi-scale features, and introduce a C2PSA attention module to improve target focusing capability:

[0030] First, the high semantic feature map P5 output by the backbone network is input into the improved spatial pyramid pooling module.

[0031] TopkSPPF is used to enhance the aggregation capability of multi-scale contextual information. For example... Figure 3 As shown, TopkSPPF consists of two 5×5 max pooling branches and one 3×3 average pooling branch. By fusing multiple local high response values ​​with the Top-k strategy through a multi-branch structure, it effectively alleviates the problem of edge information loss caused by max pooling and enhances the localization of small targets and the response capability of weak signal features.

[0032] Subsequently, a cross-stage, cross-location attention module (C2PSA) is introduced at the tail of the main branch to enhance features in the spatial dimension, and its structure is as follows: Figure 4 As shown. The C2PSA module uses a multi-head attention (MHA) mechanism to achieve adaptive weighting of spatial regions, specifically including the following calculation process:

[0033] First, grouped convolution operations are used to extract input features. Parallel generation of query, key, and value matrices:

[0034] Q = Conv Q (X),K=Conv K (X),V=Conv V (X) (2)

[0035] For each attention head i = 1, 2, ..., h, compute the scaled dot product attention:

[0036]

[0037] Where, d k For each attention head, a Softmax function is used to normalize the attention score, ensuring the model focuses on key spatial locations. Finally, the outputs of all attention heads are concatenated and fused using a 1×1 convolution to obtain the final output features.

[0038] MHA(Q,K,V)=Concat(Attention1,...,Attention h W O (4)

[0039] X'=Conv 1×1 (MHA(Q,K,V))(5)

[0040] Among them, W O The output projection weights are X′, which represents the fused features after multi-head attention. The C2PSA module enhances the long-range dependencies between different spatial locations through the above mechanism, enabling the model to effectively distinguish between targets and noise regions even in low-light and complex backgrounds, thereby improving the overall detection accuracy.

[0041] Step 3: Embed the EdgeFca module in the feature fusion stage to combine wavelet edge extraction and frequency domain weighting mechanisms, thereby enhancing feature representation under low illumination conditions, as detailed below:

[0042] like Figure 5As shown, in the feature fusion stage, this invention embeds the EdgeFca edge frequency domain joint enhancement module after downsampling of multi-scale feature maps P3 and P4. This module is used to highlight edge regions, suppress noise interference, and improve the detectability of targets in low-light scenes. Figure 6 As shown, this module consists of the following three sub-modules:

[0043] Edge extraction module:

[0044] For the input feature map Where N is the batch size, C is the number of channels, and H and W are the height and width of the feature map, respectively. Haar wavelet transform is used to discretize each channel, yielding the low-frequency approximation coefficients cA and the detail coefficients cH, cH, and cV in three directions. The edge map for each channel is calculated as follows:

[0045] E i =cH i +cV i (6)

[0046] Construct an edge tensor by overlaying all channel edge maps The original spatial dimensions are restored through upsampling to obtain the edge enhancement map E'. Then, the original input features X and E' are element-wise multiplied, and edge enhancement features are generated using the EdgeConv edge convolution module, which combines depthwise convolution and pointwise convolution.

[0047] F edge =σ(W point ·GELU(W depth ·(xe E'))) (7)

[0048] Frequency domain channel attention mechanism:

[0049] Differentiable Discrete Cosine Transform (DCT) is used to analyze F edge Perform frequency domain modeling. First, for each input channel... By predefined frequency index (u i ,v i Calculate the frequency response value f i :

[0050]

[0051] The two-dimensional DCT basis functions are defined as follows:

[0052]

[0053] In the above formula, (h,w) represents the spatial pixel position, (u i ,v iThe frequency index selected for channel i is represented by ). After extracting the frequency domain descriptor, channel attention weighting is applied through a fully connected layer (FC) and then applied to the original features:

[0054] F'=X·Sigmoid(FC([f1,...,f c ])) (10)

[0055] Dynamic residual fusion strategy:

[0056] To enhance the robustness and adaptability of feature fusion, a learnable dynamic fusion module is introduced, which performs residual weighted fusion of the original feature X and the enhanced feature F' to obtain the final output:

[0057] Y=αX+(1-α)F' (11)

[0058] Here, α represents the learnable fusion weights with a value range of [0,1], and is normalized using the Sigmoid function. This fusion strategy allows the model to adaptively choose to retain the original structural information or emphasize enhanced features under different input conditions, effectively improving overall detection performance.

[0059] Step 4: Design the DNiN decoupled detection head and complete model training and optimization, as detailed below:

[0060] This invention introduces a lightweight DNiN (Decoupled Network-in-Network) decoupled detection head at the end of the detection framework, enabling separate regression of target categories and bounding boxes, further improving detection performance and deployment efficiency. Figure 7 As shown, the DNiN detection head employs a dual-branch parallel structure to perform object classification and bounding box location prediction, respectively. Each branch uses a depthwise separable convolutional structure, constructing a pixel-level sliding perceptron through pointwise convolution to achieve efficient feature extraction and decision-making. To reduce computational complexity, the traditional single 5×5 convolution is decomposed into two 3×3 convolution operations, and the spatial relationship between classification and regression is optimized through depthwise convolution.

[0061] In the bounding box regression branch, this invention uses the relative distances {l,t,r,b} from the current position to the four bounding boxes as the regression objective, employing a distribution-based prediction-based modeling approach to reduce the dependence on precise values ​​of single points. This method uses Distribution Focal Loss (DFL) to model the bounding box regression problem, with the following form:

[0062]

[0063] in, This represents the discrete probability value predicted by the model. These are the interpolation coefficients for two discrete locations near the target value.

[0064] In the classification branch, to improve the training's ability to focus on difficult-to-classify samples, a classification loss function, Quality Focal Loss (QFL), is introduced, which takes the following form:

[0065] L QFL =-|y-σ| β ((1-y)log(1-σ)+ylog(σ)) (13)

[0066] Where σ is the predicted class score, y∈[0,1] is the joint classification and localization quality score, and β is a moderating factor that controls the attention given to difficult samples. Finally, the overall loss function is shown below:

[0067]

[0068] Where, N pos The number of positive samples is represented by λ, which is typically set to 2. GIoU represents the generalized intersection-union loss, used to optimize the overlap between the predicted and ground truth bounding boxes. This invention deploys the described DNiN detection head within a lightweight detection framework and completes training and optimization on a constructed low-light spatial image dataset. Experimental results show that this detection head structure effectively improves detection accuracy while significantly reducing model parameters and computational load, making it suitable for deployment on resource-constrained spaceborne edge platforms.

[0069] The above description is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention. Structures, devices, and operating methods not specifically described or explained in this invention are implemented according to conventional methods in the art unless otherwise specified or limited.

Images