Detection of Fiber Flaw on Pill Surface Based on Improved Deep Convolution Neural Network

Aiming at solving the problems of traditional artificial defect detection's low efficiency, strong subjectivity and heavy workload, a method based on machine vision is used to improve the detection efficiency of fiber-flaw pills. However, the fiber flaw on pill surface always has many different forms and types, which makes it difficult to detect automatically by traditional image processing method. Traditional machine vision detection technology is difficult to extract appropriate features for things with complex surface texture and fine scratch defects, and obtain satisfactory detection results. Therefore, a deep learning network based on improved VGG16 is designed for automatic detection of fiber-flaw pills. In this work, high-resolution images with pills are obtained by the independently designed pill image acquisition system. To build the data set of single pill images, the original images are processed by image processing method of inverse binarization, flood filling, and so on. Then, the S-BN-VGG (Simplified Batch Normalization VGG) model is established to distinguish fiber-flaw pills. The experimental results have shown that the recognition rate of S-BN-VGG network is 99.43%, the average detection time of S-BN-VGG network is 2.98 ms/pill, which is more accurate and faster than the traditional VGG16 model. Meanwhile, the training time of S-BN-VGG model is reduced by 68.8% compared with VGG16 model. In conclusion, the proposed S-BN-VGG network is reliable and fast, which provides an effective direction method for the on-line automatic fiber-flaw pill detection.