Lizard scale fringe structure extraction method based on deep learning and geometric topology

Abstract

The present application particularly relates to a lizard scale structure extraction method based on deep learning and geometric topology, aiming to solve the problems of large performance consumption, weak generalization ability and lack of end-to-end automatic process in traditional methods. The DETR architecture is used to automatically locate the foot region from the lizard image, the BiRefNet bidirectional reference network is used to simultaneously reconstruct the lizard scale structure and the foot body, the Douglas-Peucker algorithm is used for multi-scale polygon approximation to initialize the key points, and the iterative smoothing mechanism of local duty cycle classification is introduced to separate the main body and the adhesion impurities, the topological potential field map is constructed by mask difference, the local maximum value is extracted as the seed point, the watershed algorithm is applied to realize the optimal segmentation of the adhesion scale individual, the false detection points are removed through the area stability filtering and secondary clustering, and the geometric feature modeling is carried out. The automatic, high-precision extraction and quantitative analysis of extremely fine biological structures are realized.

Images