Defect Detection of Bottled Liquor Based on Deep Learning

In the production process of bottled liquor, due to the influence of raw material quality, processing technology and other factors, there may be various types of defects in the product that affect the product quality. Due to the variety of defects and the small size of some defects (such as liquor defects), it is difficult to detect, and manufacturers often need to invest a lot of labor costs for product quality inspection. In response to the above problems, we adopt two methods. First of all, for the dynamic liquor, we adopt a more unique image processing method, which can not only retain the original information and location information of the picture, but also keep the difference information as a guide. Secondly ,our propose a defect detection algorithm for bottled liquor based on deep learning, which contains many structures that can improve the defect detection performance, it has many advantages: 1) ROI Align replaces ROI Pooling, eliminating quantization error, 2) FPN can greatly improve the detection performance of small objects without increasing the calculation of the original structure, 3) The cascade algorithm makes the output distribution of each stage of detectors conducive to training a higher quality detector in the next stage with a higher IoU threshold, 4) Deformable convolution network(DCN) better fits the target of bottled liquor defects, meanwhile, some other techniques are also used to improve the accuracy. Experiments show that the above method can greatly improve the accuracy, and we also test the time requirement to ensure that the accuracy of the model decreases slightly while the model has a faster detection speed.