Anchor-frame-free instance segmentation method for Mars meteorite crater

Abstract

The invention discloses an anchor-frame-free instance segmentation method for a Mars meteorite crater, and relates to the technical field of computer vision and deep space exploration. According to the method, the Swin Transform is adopted as a backbone network to realize context sensing feature extraction, and the problem of meteorite crater misrecognition caused by rugged Mars landforms is solved; the semantic dislocation problem in feature fusion is solved through a semantic detail enhancement up-sampling module, so that the accuracy of meteorite crater boundary detection is improved; multi-scale feature fusion is completed in combination with a feature pyramid network; and finally, outputting category information and polar coordinate morphological parameters of the meteorite craters through the anchor-frame-free instance segmentation head. The method does not need to generate a candidate area, greatly improves the calculation efficiency, effectively adapts to the multi-scale characteristics and circular geometric characteristics of the Mars meteorite crater, solves the problems of meteorite crater error detection and missing detection and insufficient contour prediction precision, meets the resource constraint of an edge calculation scene of a Mars patroller, and improves the calculation efficiency. The method can be directly applied to an autonomous navigation and path planning system of the patroller.

Description

Technical Field

[0001] This invention relates to the fields of computer vision, deep space exploration technology and artificial intelligence, and specifically to a frameless instance segmentation method for Martian craters. Background Technology

[0002] Accurate instance segmentation of Martian craters is a core technological support for the autonomous navigation and path planning of Mars rovers, directly impacting the safety and reliability of deep space exploration missions. Existing instance segmentation methods are mainly divided into two-stage methods and single-stage anchor-free methods:

[0003] The two-stage approach requires generating candidate regions through a region proposal network and then performing instance segmentation through operations such as ROI Align. However, it suffers from high computational complexity and slow inference speed, making it difficult to meet the resource constraints of edge computing scenarios for Mars rovers. Furthermore, the ROI pooling operation in the two-stage approach is prone to causing loss of small-scale crater target information, posing a risk of missing small targets.

[0004] Single-stage anchor-free methods eliminate the need for candidate region generation, resulting in higher computational efficiency and adaptability to small target detection requirements through dense prediction mechanisms. However, existing anchor-free methods have shortcomings in their feature upsampling modules: traditional bilinear interpolation and deconvolution operators lack content awareness and have limited receptive fields; the CARAFE content-aware upsampling algorithm relies solely on low-resolution deep features, ignoring detailed information in high-resolution shallow features, leading to semantic misalignment in the upsampled feature maps and affecting multi-scale feature fusion and segmentation accuracy.

[0005] Therefore, there is an urgent need for an instance segmentation method that balances computational efficiency and segmentation accuracy, and is adapted to the multi-scale characteristics of Martian craters and edge computing scenarios. Summary of the Invention

[0006] 1. Purpose of the invention

[0007] This invention aims to solve the problems of low computational efficiency, missed detection of small targets, semantic misalignment, and insufficient contour prediction accuracy in existing Martian crater instance segmentation methods, and provides an anchor-free detection method and system to meet the real-time and reliability requirements of Mars rover autonomous navigation.

[0008] 2. Technical Solution

[0009] The core technical solution of this invention is as follows:

[0010] 2.1 Overall Architecture Design

[0011] The anchorless detection framework of the present invention comprises four parts: an image encoder, a semantic detail enhancement upsampling module, a feature fusion module, and an anchorless instance segmentation head, achieving end-to-end optimization from feature extraction to instance segmentation.

[0012] 2.2 Context-Aware Feature Extraction

[0013] Using the Swin Transformer as the backbone network, a hierarchical design is employed to generate multi-resolution feature maps, preserving the spatial details and semantic information of craters at different scales. A shift-window mechanism is used to alternately perform local attention calculations and global information transfer, reducing computational complexity while strengthening the spatial correlation of crater edge points through a self-attention mechanism, thus adapting to the geometric characteristics of circular targets.

[0014] 2.3 Semantic Detail Enhancement Upsampling Module

[0015] This module introduces detailed information from high-resolution shallow features based on the CARAFE algorithm to alleviate semantic misalignment problems. It includes:

[0016] The kernel prediction module consists of a channel compressor, a content encoder, and a kernel normalizer. The channel compressor reduces computational overhead by compressing the number of channels through 1×1 convolutions; the content encoder extracts local contextual information to generate basic kernel features; and the kernel normalizer ensures that the sum of kernel weights is 1 through Softmax.

[0017] Kernel calculation module: Uses the recombination kernel to perform weighted aggregation on local regions of input features to complete content-aware upsampling.

[0018] Two-stage detail enhancement: The first stage breaks down kernel parameter generation into semantic expansion and detail-guided recombination. The second stage enhances the injection of detailed information by summing shallow kernel features and combined kernel features.

[0019] 2.4 Multi-scale feature fusion and instance segmentation

[0020] Adaptive fusion of multi-scale features is achieved by fusing upsampled features and shallow features through a feature pyramid network. The instance segmentation head adopts a shared parameter design, with the category prediction branch outputting pixel category probabilities and the regression branch outputting polar centrality and polar distance.

[0021] 2.5 Collaborative Optimization Loss Function

[0022] The design incorporates a collaborative optimization objective with four types of losses to address class imbalance, distance regression accuracy, contour representation error, and centrality prediction optimization, ensuring accurate segmentation of crater instances by the model.

[0023] 3. Beneficial effects

[0024] The present invention has the following beneficial effects:

[0025] High computational efficiency: It adopts an anchorless framework structure, which eliminates the candidate region generation and ROI-related operations. Combined with the shift window mechanism of SwinTransformer, it greatly reduces computational complexity and meets the resource constraints of edge computing scenarios.

[0026] Strong multi-scale adaptability: Through feature pyramid network and dense prediction mechanism, it can naturally adapt to Martian craters of different scales and effectively avoid missing small targets.

[0027] High segmentation accuracy: The semantic detail enhancement upsampling module solves the semantic misalignment problem of feature upsampling, and collaboratively optimizes the loss function to simultaneously optimize classification, regression and contour representation, thereby improving the accuracy of crater contour prediction;

[0028] Highly practical: It can be directly integrated into the autonomous navigation and path planning system of Mars rovers, providing technical support for deep space exploration missions. Attached Figure Description

[0029] Figure 1 This is a diagram of the semantic detail enhancement upsampling structure of the present invention.

[0030] Figure 2 This is a structural diagram of the core computing module of the present invention.

[0031] Figure 3 The segmentation result of the example segmentation head of the present invention is obtained by connecting the predicted center of each target and the length of 36 rays emitted from the center.

[0032] Figure 4 This paper compares the semantic detail enhancement upsampling method of the present invention with the nearest neighbor interpolation upsampling and CARAFE upsampling methods. The first column is the input Mars image, the second column is the label image, the third column is the feature map obtained by upsampling in the present invention, and the fourth and fifth columns are the feature maps obtained by nearest neighbor interpolation upsampling and CARAFE upsampling, respectively. The comparison results show that the semantic detail enhancement upsampling method of the present invention can fully integrate deep semantic features and shallow detail features. Compared with other upsampling methods, it can effectively remove noise and has excellent detail processing capabilities. Detailed Implementation

[0033] 1. Image preprocessing

[0034] Acquire raw images of the Martian surface, which can be taken by the optical camera onboard the Mars rover. Standardize the resolution of the raw images, adjusting the image size to 512×512 pixels; then normalize the pixel values ​​using Z-Score normalization to achieve a mean of 0 and a variance of 1, resulting in a standard input image.

[0035] 2. Context-aware feature extraction

[0036] The standard input image is fed into the Swing Transformer backbone network, which has a four-layer structure. Each layer includes multiple shift-window attention modules and a feedforward network. The first layer outputs a feature map with a resolution of 128×128, the second layer outputs a feature map of 64×64, the third layer outputs a feature map of 32×32, and the fourth layer outputs a feature map of 16×16, forming a multi-resolution feature map set.

[0037] Each layer's shift-window attention module first divides the feature map into non-overlapping windows and calculates local attention within each window; then, it shifts the window position to perform global information transfer across windows, ensuring the global modeling capability of the feature map. The self-attention mechanism strengthens the spatial correlation of crater edge points by calculating attention scores between queries and key values, adapting to the characteristics of circular targets.

[0038] 3. Semantic detail enhancement upsampling

[0039] The feature maps of layers 1 and 2, layers 2 and 3, and layers 3 and 4 are input into the semantic detail enhancement upsampling module to obtain upsampled feature maps with sizes of 128×128, 64×64, and 32×32.

[0040] 4. Multi-scale feature fusion

[0041] The upsampled feature maps are input into the feature pyramid network along with the original feature maps of layers 1, 2, and 3, respectively. Through lateral connections and top-down path fusion, a fused feature map containing multi-scale semantic and detailed information is generated.

[0042] 5. Instance Segmentation and Model Optimization

[0043] The instance segmentation head processes the fused feature map: the class prediction branch uses a 3×3 convolution and a sigmoid activation function to output the crater / background class probability for each pixel; the regression branch outputs polar centrality and polar distance, predicting 36 epipolar lines for each pixel. During training, a stochastic gradient descent (SGD) optimizer is used with an initial learning rate of 0.001, iteratively training until the loss function converges.

Claims

1. A method for anchorless instance segmentation of Martian craters, characterized in that, Includes the following steps: Step 1: Image input and preprocessing. Obtain the original image of the Martian surface, perform resolution standardization and pixel normalization processing to obtain the standard image for the model input. Step 2: Context-aware feature extraction. The standard image is input into the backbone network based on Swin Transformer. Through hierarchical design and shift window mechanism, local attention calculation within the window and global information transfer across the window are performed alternately to output a multi-resolution feature map set. The feature map set contains spatial details and semantic information of craters at different scales. Step 3: Semantic detail enhancement upsampling. The low-resolution deep feature map output in Step 2 is upsampled using the semantic detail enhancement upsampling module. The semantic detail enhancement upsampling module includes a kernel prediction module and a kernel calculation module. The kernel prediction module generates a reconstructed kernel that fuses shallow detail information and deep semantic information through a channel compressor, a content encoder, and a kernel normalizer. The kernel calculation module uses the reconstructed kernel to perform weighted aggregation on the input features and outputs a semantically aligned high-resolution feature map. Step 4: Multi-scale feature fusion. The upsampled feature map output from Step 3 and the high-resolution shallow feature map output from Step 2 are input into the feature pyramid network to achieve adaptive fusion of multi-scale features and semantic consistency optimization. Step 5: Anchor-free instance segmentation. Using the center of the instance as the origin, emit m uniformly distributed dense rays. The length of each ray corresponds to the distance from the origin to the instance boundary, forming a distance set D = {d1, d2, ..., d...}. m The network needs to simultaneously complete the initial point classification and dense distance regression, and generate the polygonal outline of the instance through polar coordinate parameter inversion. The fused feature map is processed by the instance segmentation head with shared parameters. The category prediction branch outputs the category probability of each pixel, and the regression branch outputs the polar centrality and polar distance of each pixel. The model parameters are optimized by combining the preset loss function, and finally the instance segmentation result of Martian craters is output.

2. The method according to claim 1, characterized in that, The kernel prediction process of the semantic detail enhancement upsampling module includes two-stage detail enhancement: Phase 1: The content encoder expands the kernel size of the low-resolution deep feature map to obtain the basic kernel features; the kernel prediction module generates a detail information reconstructed kernel using the high-resolution shallow feature map; the basic kernel features and the detail information reconstructed kernel are input into the kernel calculation module to obtain the combined kernel features that fuse semantics and preliminary details. The second stage involves summing the original kernel features corresponding to the high-resolution shallow feature map with the combined kernel features, and then performing Softmax normalization processing to obtain the final semantic information reconstructing kernel.

3. The method according to claim 1, characterized in that, The operation process of the kernel computing module includes: For each source location in the input feature map, extract a k×k local region centered at that location, where k is the size of the reconstruction kernel; expand the σ×σ square region corresponding to the reconstruction kernel location along the channel dimension to obtain σ 2 The input feature map is multiplied by its corresponding k×k recombination kernel weights, where σ is the upsampling factor. σ is obtained by multiplying each channel of the input feature map by its corresponding recombination kernel weight. 2 The feature vectors are concatenated to form an upsampled feature of size σ×σ; the upsampled features at all positions are concatenated to obtain a high-resolution feature map of size σH×σW, where H and W are the height and width of the input feature map.

4. The method according to claim 1, characterized in that, The loss function of the instance segmentation head is a collaborative optimization objective, expressed as follows: Loss=λ cls L cls +λ reg L reg +λ poly L poly +λ sre L sre (1) The classification loss L cls Focal Loss is used to mitigate the class imbalance problem; regression loss L reg We continue to use PolarIoU Loss to optimize the accuracy of ray distance regression. Where d i The distance to the truth value. For predicting distance; L poly Using IoU loss, the predicted distance is transformed into a polygon pixel mask. The IoU loss between the pixel mask and the instance label is calculated to directly quantify the actual overlap between the predicted polygon and the instance contour; L sre The centrality prediction head is optimized using cross-entropy loss and Polar Centerness as the supervisory signal.

Images