Detection Of Appearance Defects Of Cigarettes Based On Improved SSD

Tobacco is one of the important economic crops in my country.Yunnan Province,as the "Kingdom of Tobacco",has ranked first in the country for many years in terms of tobacco quality,output,sales volume and export earnings.Important support and main source of tax revenue.As the most common product of tobacco,cigarettes have always been the focus of the development of the tobacco industry.Although the current high-speed production line of cigarettes can meet the demand for output,it will inevitably increase the difficulty of quality inspection of cigarettes.In the quality inspection,the appearance defects of cigarettes will directly affect the quality of cigarettes.If a large number of defective products enter the market,it will not only reduce the satisfaction of consumers,but also affect the value and economic benefits of the brand.With the development of industrial technology,manual-based quality inspection methods have fallen behind,and computer vision technology is widely used in the quality inspection of various products.At present,the image processing technology combined with deep learning,with the help of optical equipment such as industrial high-speed cameras and cameras,can effectively detect the defects of products on the industrial production line,which provides a new idea for the defect detection of cigarettes.However,when the existing methods are used in actual production lines,there are still some shortcomings in terms of speed and accuracy.Based on the above problems,this paper studies the real-time appearance defect detection of cigarettes in combination with the actual needs of enterprises.The specific work is as follows:1.Aiming at the problem that the original SSD network model has poor accuracy in detecting small targets,and considering the characteristics of the obtained cigarette appearance defect data and the small defect scale,an improved SSD model is proposed to improve the appearance of cigarettes.Defect Detection Network.(1)In this paper,the SSD network model is used as the basic network architecture,the ResNet50 network with better effect is used as the new feature extraction network,and a series of feature layer structures are added on this basis;(2)In order to improve the network’s performance for small targets Detection accuracy,based on the idea of feature fusion,adding pyramid convolution to increase the receptive field of the convolution layer and the utilization of shallow semantic information;(3)In order to make the model more focused on the target information,the feature attention mechanism module ECANet is added,and the experiment proves that the improved method is effective for improving the accuracy of the model;2.After improving the structure of the network model,the activation function and loss function during model training are also optimized:(1)The original ReLU neuron is relatively "fragile",and it flows through a large gradient and updates the parameters.After that,you will no longer be able to activate any data.Therefore,in view of this problem,a new activation function P-ReLU is compared and introduced,which can better alleviate the limitation that the original activation function is easy to "die" in training;(2)The number of categories in the data set based on the appearance of cigarettes used is unbalanced(3)Optimize the optimizer of the model to further improve the detection accuracy of cigarette appearance defects and enhance the robustness of the network model.Through experiments,the effectiveness of the algorithm proposed in this paper is verified in the detection of cigarette appearance defects.