Mobile Phone Surface Defects Recognition Method Based on Multi-scale Convolution Neural Networks

To address the problem that mobile phone surface defects are difficult to be identified accurately, an approach for mobile surface defects recognition called SL-MSCNN was proposed, which combined the Sobel operator, logistic loss function (LLF) and multi-scale convolutional neural networks (MSCNN). First, a neighborhood feature enhancement method based on Sobel was constructed, which can exclude the interference of unrelated factors of image, such as lighting and shadow. Second, a defect recognition method based on MSCNN was designed, which can improve the recognition accuracy of mobile phone surface by obtaining multi-scale information of mobile phone surface image. Meanwhile, LLF was introduced, which speeded up the detection speed of training by reducing the probability of gradient disappearance. Results show that the proposed defect recognition method, compared with other methods, can improve the practicability of the actual process in terms of recognition accuracy and efficiency.