A Novel Superpixel Segmentation Method Based on Adaptive Seed Expansion Random Walk Algorithm Toward Complex Scene Images Processing

Superpixels facilitate image complexity reduction by over-segmenting it into compact atomic regions that align with object boundaries. However, most superpixel segmentation algorithms struggle to achieve an ideal balance between boundary adherence and computational efficiency, especially in complex scene images. In this article, a novel superpixel segmentation method called adaptive seed expansion random walk (AERW) algorithm is proposed based on the random walk (RW) model. A multifeature adaptive fusion model is first proposed, which determines different feature weights based on image content to achieve new feature representation. Seed initialization and seed expansion are then implemented based on the proposed multifeature adaptive fusion model. The seeds are evenly distributed and avoid falling on the object boundary to ensure the shape and boundary adherence of superpixel through the proposed seed initialization. The traditional RW graph model is improved by seed expansion. The computation of edge weights in the graph model and the vector dimension in label membership probability calculation are reduced, which improves operating efficiency and reduces node label assignment errors. The experimental validation carried out with public datasets and actual mechanical machining photographs, demonstrates that the proposed AERW presents lower computational cost and better boundary adherence than existing models in the complex scene image. The proposed AERW algorithm is expected to provide enhanced robustness for preprocessing tasks in industrial vision measurement and inspection that necessitate high precision.