Research On Appearance Defect Detection Of Cigarette Capsules Based On Deep Learning

Cigarette capsules are the gelatinous spheres that provide various flavors for cigarettes in the tobacco industry.Quality control of cigarette capsules is an important part of cigarette production;with the advent of the "smart manufacturing"era,how to achieve industrial intelligence is an important research topic.Cigarette capsules are prone to various defects in the production process,and these defects can reduce the quality of cigarette products.At present,cigarette manufacturing enterprises mainly use the method of manual identification of defects,which has low efficiency and high cost.Machine vision technology can detect appearance defects of cigarette capsules,whereas traditional image processing-based methods have drawbacks such as limited image feature expression capability and poor defect adaptation.This thesis investigates the deep learning-based appearance defect detection technology of cigarette capsules in order to improve the robustness and automation of cigarette capsule defect detection.The main work is as follows:(1)A cigarette capsule defect detection device is designed and built in response to the actual industry demands.It specifically includes the design and selection of key equipment for modules such as supply transmission,light source,image acquisition,image recognition,and product processing.The image dataset of cigarette capsules was then constructed using this inspection device,and the data was pre-processed.(2)For the problems of poor image feature representation and difficult model optimization,a one-dimensional residual network model based on multi-feature fusion(1D ResNet_mix)is proposed.Firstly,multiple image features such as shape,texture,and color are extracted and fused into a global feature vector.Specially,a novel image feature descriptor named Histogram of Edge Directions(HED)is proposed to characterize image texture.Subsequently,a 1D deep residual neural network for feature classification is designed.Experiments show that HED descriptor can accurately and efficiently depict image defect information,and the 1D ResNet_mix not only can effectively detect defects,but also has the advantages of low complexity and a small number of parameters,which enable the model to be trained and migrated quickly with limited dataset and hardware resources.(3)To address the limitations of manual image features,unbalanced data,cumbersome labeling of defect samples,and the poor applicability of scenarios with supervised methods,a model named CMA-Net is proposed for the appearance defect detection of cigarette capsules.Firstly,the model learns self-supervised representations via a proxy classification task between normal data and the ones augmented by CutPaste.Then,A multi-scale attention network is proposed to improve the model’s ability to focus on and learn from defects at different scales.Experiments show that multi-scale attention can improve the model’s ability to identify defects,CMA-Net has higher accuracy and robustness in the appearance defect detection of cigarette capsules when compared to current advanced self-supervised learning algorithms.This dissertation studies the application of deep learning in the appearance defect detection of cigarette capsules,improving the limitations of traditional image processing methods in defect detection.A system for appearance defect detection of cigarette capsules is designed,and two approaches with real-world industrial applicability are proposed.