Intelligent Impact Crater Detection on the Surface of Mars Based on YOLOv7

Craters are the most typical and widespread topographic features and geological structures on the lunar surface. Real-time crater detection can be used to achieve missions such as autonomous landing, spacecraft and lunar rover navigation. However, due to the complex data distribution, detecting craters is a difficult task, and there are also many small craters in a crater picture. To solve this problem, this paper uses YOLOv7 algorithm, which is currently used in the area of objection detection and use a CNN building block called SPD-Conv to take place the stepwise convolutional layer and each pooling layer. Then a new layer of small target detection is added to the original neural network architecture, and the Wasserstein distance metric is used to help the IOU metric originally used in YOLOv7, reducing the error in small object recognition. Finally, BiFPN is added to YOLOv7 to fuse more scale features. In addition, this study recreated a new dataset to detect small craters. This paper conducted extensive experiments based on this new dataset, and the results have shown a good performance by using this kind of network, especially in detecting the small craters. The recall and precision of this method can get 86.45% and 88.40% respectively, which is better than most detection algorithms.