Research On Medicine Bottle Defect Detection Algorithm Based On Deep Learning And Zernike Moment

The special effects of medicines on health put forward higher requirements on the quality of medicine bottles,which must be inspected after production and before being used for packing.Manual inspection is subject to missed detection,mischeck,slow speed.And long time gazing in front of the light source will inevitably cause damage to the eyes.Machine vision can replace human eyes for detection,but traditional image processing algorithm is still used to solve the appearance defects of bottle size and simple parts in the current practical application.The current algorithms suffer from low precision,slow speed and poor reliability,and cannot be applied to the industry when there are various and complex defects in medicine bottles.In this paper,a unique optical image acquisition system is designed to collect the image of the bottle mouth,body and bottom in a comprehensive way,aiming at more than 30 kinds of defects such as inclined bottom,bubble,shoulder stone,twist neck and uneven mouth.The in-depth study has been conducted on the detection of the above defects,and a medicine bottle defect detection algorithm based on deep learning and Zernike Moment is proposed.The main research contents and contributions are as follows:(1)A Semantic-Pixel Fusion detection(SPFD)algorithm is proposed for producers to adjust their criteria for the tolerance of defects.The algorithm classifies and locates defects through semantic and region-level features,and then integrates traditional algorithms to evaluate the length and area of tolerable defects at the pixel level,so that the evaluation criteria can be adjustable.(2)Siamese Defect Classification Network(SDCN)is proposed to realize unknown defect detection.The network is trained based on positive samples and known defect samples to obtain the model.When the unknown defect samples are input into the model,the model can detect significant differences between the defect samples and the positive samples.The network focuses on the difference between the sample to be detected and the positive sample,rather than the specific characteristics of a defect,thus improving the generalization ability of the algorithm.(3)A defect classification method based on Zernike moment and a GPU acceleration algorithm with merged kernel are proposed for the classification of large defects at the bottle mouth and bottom.The 20-order Zernike moment is used as the image feature,and the feature is carried out dimensional mapping and principal component analysis,followed by SVM training to obtain the classification model.In addition,the combination of cores in the GPU improves the computing speed.The proposed algorithm model wa trained using 4594 self-made data sets.In the factory test,the algorithm was tested on the experimental platform with CPU i7-10700 and GPU RTX 1660.The test results show that the algorithm can detect multiple and multi-level defects on the whole bottle surface,and the comprehensive accuracy is more than 99%,which greatly meets the actual requirements of manufacturers.