Research on region segmentation algorithm oriented to solar cell surface image

With the reducing of disposable energy and the worsening of environmental pollution，solar energy，a sustainable green energy, has gained widely concerned. Similarly, solar panels as the carrier of solar power generation has attracted much attention. The quality of solar cells is an important factor affecting the efficiency of power generation, and an indispensable part of the power generation process. The segmentation algorithm for solar cell surface image segmentation has high segmentation complexity and the number of targets that needs to be manually intervened and the accuracy of semantic expression that is inaccurate. Concerning these three issues, This document of this project is to explore an image segmentation algorithm, which has the ability to learn autonomously, has high boundary fit, and the low computational complexity.The main research contents are as follows: Firstly, aiming at problem of manual intervention in superpixel segmentation, the multi-task circular convolutional neural network (Multi-mask RCNN) is proposed to train and learn the sample data with weak annotation information to achieve accurate and fast segmentation. Secondly, for low boundary fit for segmentation, it is proposed to add texture information as optimisation information to the evaluation function to guide the segmentation and ensure the integrity of edge segmentation. Finally, to solve high computing complexity, the segmented region is considered as the vertex of the graph of spectral clustering, and spectral clustering algorithm is applied to cluster the regions with similar features and semantic relatedness to achieve the accurate segmentation of the surface images of the solar panels.