Research On Bubble Detection Technology Of Composite Film Based On Deep Learning

Intelligent manufacturing promotes the rapid development of the packaging and printing industry.In the process of rapid development,it is even more critical to control the quality of products.The production process of flexible packaging is complicated.Bubble defects often appear in the film lamination process,which is also the key to the quality of packaging bags.At present,flexible packaging is widely used in food,medicine,cosmetics and other fields.Unqualified packaging bags will affect packaging.The shelf life of the product affects human health.At present,the detection of bubble defects in the compounding process mainly relies on workers,which is inefficient and often misses detection,which affects subsequent production.This thesis proposes a method for detecting bubble defects in composite films based on deep learning,builds an improved YOLOv5 algorithm model,and extracts features of bubbles to achieve high-precision detection of bubbles.The first is the construction of the composite film bubble data set.Collect images of composite film bubbles in flexible packaging manufacturers,and preprocess them,and reduce the size of the images by segmentation to make them more suitable for the input of the YOLOv5 network;divide the bubbles into three categories: cluster,big,and small,to solve the problem The problem of difficult extraction of bubble features caused by the different sizes and shapes of bubbles is solved;the performance of the model is improved by using data enhancement.After experimental testing,the m AP@0.5 of the model after data enhancement has increased by 19.7%.Three data sets were mainly constructed.The first data set is the full data set,with a total of 22589 pictures.The second is a sub-data set.1500 pictures are randomly selected from all data sets.The third data set is the original data set,a total of 3227 pictures,the above three data sets are randomly divided into 3 parts according to the ratio of 6:2:2,which are used as the training set,verification set and test set in turn.The second is the improvement based on the YOLOv5 algorithm,which is mainly divided into three aspects.First,the optimization of the above data sets by the SGD optimizer and the Adam optimizer is compared.The experimental results show that the SGD optimizer is more suitable for the optimization of the compound film bubble data set.m AP@0.5 reached 94.3%;the second is based on the improvement of the attention mechanism,introducing four attention mechanism modules: SE,CBAM,ECA,and CA,and adding them to the front of the SPPF layer and the C3 structure of the Backbone module of the YOLOv5 network,respectively.The experimental results show that adding the CA attention mechanism to the C3 structure improves the performance of the model the best,with m AP@0.5 reaching 95.1%;the third is the lightweight model,introducing Ghost Net,and adding the attention mechanism to the lightweight network The model fusion of YOLOv5-C3CA-Ghostnet-D algorithm model is obtained,and the bubble detection obtained by this algorithm is obtained by setting 5 groups of experiments on the sub-data set to test the influence of the position of the Ghost Net and C3 CA attention mechanism in the Backbone module on the results.The model is reduced from the original 13.7MB to 11.7MB,and the weight reduction effect is obvious.Finally,through the above experiments and research,it is determined that YOLOv5-C3CA-Ghostnet-D is an algorithm model suitable for composite film bubble detection.Using the full data set to train the model,the accuracy rate of the composite film bubble detection model was finally obtained at 91.8%,the recall rate was 90.1%,and the average precision was 95.3%.The performance test shows that the bubbles on the picture can be accurately identified with a high degree of confidence.The average detection speed FPS is 84f/s,which can meet the needs of actual production on-line detection,and has certain reference significance for the detection of composite film bubbles.