Research On Key Technologies Of Photoelectric Automatic Recognition And Classification Of Cigarettes

Tobacco sorting or tobacco review is an important part of the tobacco distribution system.Studying the automatic identification and classification technology of tobacco accelerates the automated process of tobacco enterprises,frees laborers from repeated and tedious manual sorting lines,and improves the economic benefits of tobacco enterprises.It has important significance.The use of machine vision instead of the human eye,and the use of digital image processing technology to replace the human brain,are a major trend in the field of industrial detection and identification.The accuracy of the automatic identification and classification of cigarettes directly determines the feasibility of the automated cigarette sorting and review scheme.Therefore,it is necessary to study the classification and recognition algorithm of cigarettes with high accuracy,high speed and high robustness.There are two main ideas for the identification and classification of objects.The first one locates the target position firstly,then separates the target,extracts the features of the target and compares them with the similarity of the existing category criteria.The other uses fast-developing deep learning technology recently.By training network structure through a large number of data,images are input into the network to obtain classification results.The main research contents of this thesis are as follows:Due to the plastic packaging on the surface,the use of natural light can cause differences in brightness and contrast.Rather,it is difficult to avoid the reflection of the surface with conventional light sources,and the cigarettes move quickly on the conveyor belt.This thesis builds a suitable optical system and image acquisition system through experiments,which is very important for further identification and classification.Aiming at the detection requirements,this thesis studies a classification method based on feature extraction and matching.First,the foreground and background regions are segmented by the image preprocessing algorithm,and then the position of the target is determined by contour detection,and the target is intercepted by affine transformation.For the accretive cigarettes,this thesis proposes a concave point detection algorithm based on contour and convex hull,which can quickly segment the target of adhesion.Then,for the target that has been segmented,some common features such as color,edge and texture are analyzed and extracted.Finally,a multi-feature weighted matching algorithm is established.Experiments showed that the recognition accuracy of the algorithm reaches 99.53%.Aiming at the detection requirements,this thesis studies a classification method based on SSD(Single Shot MultiBox Detector)model,and the application of deep learning technology in the direction of target detection.Firstly,the data sets are produced by a large number of pictures from the industrial scene,and then we train the cigarette classification and recognition network using the SSD algorithm.Experiments show that the algorithm has good robustness.By comparing the two methods,the feature matching algorithm is simpler and more efficient in training or modeling.In terms of accuracy,the SSD algorithm has a higher recognition rate on the test samples,and as the sample size increases,it can be further improved.