A self-reference scheme based on structure-texture decomposition for crack defect detection with electroluminescence images

Surface defect detection based on machine vision has drawn much attention today. Traditional methods aim at uniform repetitive texture, thus can rarely handle inhomogeneous texture surfaces like solar cells'. Therefore, a self-reference scheme based on the decomposition of structural-texture is introduced here to observe solar cell's surface cracks under electroluminescence (EL) images. Firstly, the structure-texture decomposition of the original image is carried out, and the $L_{0}$ gradient minimization and the relative total variational operation are carried out on the structural component and the textural component respectively. It turns out that the small amplitude gradient information in the structural map is removed and the crack details are preserved in the textural map. Then, the discrete wavelet transform is used to process the structural component and the textural component, and a self-reference image is obtained by combination. Through finding an appropriate radius in the spectrogram of self-reference image and setting the frequency domain inside the selected circular area to zero, we can finally acquire the precise location of the defect. The proposed method has been proved of high efficiency from a large set of tests of a real production line.