Multi-level Efficient Perception Network for Grain's Edge Detection of Cross-polarized Petrographic Images.

The edge detection of mineral grains in petrographic images is the first step of the analysis of petrographic images, which provides cues about the grain's size, shape, and composition. The challenge of automatic methods lies in the ambiguous edges of adjacent grains, as well as the various colors and intensities of grain under different circumstances. In this paper, a multi-level efficient perception network (MLEP) for the grain edge detection of cross-polarized petrographic images is proposed to address the above problem. The multi-angle inputs fusion block (MAIF) takes advantage of sequential petrographic images to generate edge-enhanced features, followed by the modified EfficientNetV2 and the proposed BiDecoder to obtain an affluent and hierarchal representation of edge features. The cross-polarized petrographic image datasets, named CPPID, are generated and carefully annotated. The proposed MLEP is tested on CPPID and achieves 0.888 F1-scores. Experimental results demonstrate the effectiveness of the proposed model, which outperforms four classical CNN-based edge detection models by a large margin.