Semantic Segmentation Network for Small-Body Rock Based on Improved DeepLabv3+

To address the issues of complex calculations, large memory consumption, and insufficient segmentation of small rock fragments and edge detection in traditional rock image segmentation algorithms, we propose an improved small-body rock segmentation algorithm based on DeepLabv3+. To cut down on computations and use, this method leverages the lighter MobileNetV2 network as its backbone. The HDC module is used in place of the original dilated convolution in ASPP in order to mitigate the gridding impact. The network's ability to fuse and extract features at various sizes is further improved by the addition of the Coordinate Attention (CA) and Channel Feature Pyramid (CFP) modules. The suggested approach obtains a 95.88% accuracy rate for small-body rock picture segmentation, with a MPA of 80.28% and a MIOU of 71.43%, according to experiments conducted on simulated datasets.