Abstract:By machine vision and machine learning algorithms for traditional low existed in the recognition of the wafer cracked detection precision, the recognition rate is poor, and testing takes a long time of problem, a new detection method was proposed. The optimized neural network to detect targets in the images of the single depth method of single shot multiBox detector (SSD), feature extraction of SSD combined with dense connectivity convolution network (DenseNet), the defection difficulty of extracting crack of less than 0.1 mm was solved in the original network. Through the experiment, the optimized SSD detection algorithm improved the detection accuracy of crack under 0.01 mm by 22% as compared with the traditional texture filtering and state vector machine (SVM) classification detection algorithm. The detection accuracy is 6% higher than the nonoptimized SSD algorithm. The validity of this method is proved.