LAUSR

The detection of dislocation defects in polysilicon wafers helps to improve the power generation efficiency and service life of solar cells. However, dislocation defect detection is challenging due to similarity of morphology and intensity between the non-uniform random texture background and dislocation defect regions. In this paper, a novel and robust Multi-scale Feature Saliency Map (MFSM) is proposed to segment dislocation defects accurately. In order to highlight the dislocation area and weaken the background, we employ the Parameter-optimized Atmospheric Scattering Model (PASM) to enhance image contrast and preserve dislocation defect region information. Then, the multi-scale gradient feature is employed to obtain the multi-scale feature saliency map including all possible contours from the enhanced image. Furthermore, the watershed transform is employed to remove pseudo-defective regions arcs in MFSM. Finally, super hierarchical region tree is used to rank the likelihood of dislocation contours to obtain accurate dislocation area. The experimental results show that the proposed method can effectively segment dislocation defects and have good adaptability and robustness to complex background.