Wine Bottle Defect Detection Based On Improved Cascade R-CNN

Defect detection is in great demand in many industrial fields such as textiles,glass products,steel,road traffic,chip,etc.Traditional defect detection methods include artificial method and machine vision-based method.However,the former has the shortage of low efficiency and slow speed,while the latter suffers from poor robustness and relies on some artificially designed features.Recently,with the improvement of computing power and the rapid development of deep neural network,deep learning,especially object detection,has gradually been one of the research hotspots for defect detection.Cascade R-CNN is a new two-stage object detection algorithm with good performance.Therefore,based on Cascade R-CNN and its improvement,the winebottle defects were studied.The main work and conclusions are as follows:(1)A method of k-means++ clustered by using the anchor aspect ratio is proposed to design the anchor shape in RPN network.Compared with the previous method of obtaining the anchor shape based on data analysis,the detection results of the new method is better.(2)In order to solve the problems of noise and sample imbalance in the data set,a pre-processing method combining picture clipping,small target oversampling and defect splicing is proposed.The pre-processing method is finally determined by experiments of different combinations.After using the pre-processing method to process the data set of the winebottle,the score of m AP was improved by 2.6%.(3)According to the defect characteristics of winebottle,the Cascade R-CNN network structure was improved and verified by experiment.Specifically,the backbone network from ResNet-50 pre-training network with only ordinary convolution is replaced by ResNet-50 pre-training network with deformable convolution v2.The balanced feature pyramid structure is used in FPN to enhance the semantic information of each layer feature.The Guided Anchoring structure is used to replace the RPN to extract the region of interest with higher quality.Focal Loss is used in FPN classification error to solve the unbalanced sampling problem of regions of interest.In the R-CNN network stage,an improved OHEM algorithm is used to solve the sample imbalance problem.The soft-NMS is used to replace the NMS so that the possible overlapping defect targets in the sample were retained.Although the detection speed decreased slightly,the overall m AP after using the improved Cascade R-CNN network for detection was 2.9% higher than that after using data set pre-processing only.