Research On Detection Method Of Tobacco Impurities Based On Machine Vision

With the improvement of computer performance and computing power,it has become an important research direction to apply machine vision technology with the advantages of convenience,accuracy,rapidity and intelligence in various fields of industrial production.Impurity detection and rejection in tobacco production is an important part of ensuring tobacco product quality.Applying machine vision related technology to the task of tobacco rejection can effectively alleviate the reliance on manual labor in the production process,and improve the automation level of product production while reducing labor costs.Tobacco de-mixing is all about detection.In recent years,researchers have studied impurity detection methods during product production,but there is still much room for improvement in the study of visible light-based tobacco impurity detection methods.In this paper,we use machine vision related technology,combined with deep learning to study the tobacco impurity detection algorithm,which can provide a referenceable idea for impurity detection in the tobacco production process.We collected visible images from tobacco industrial production lines and produced a dataset specifically for tobacco impurity detection task by means of crop and data enhancement,and based on this,we proposed the FARLut model and FSLut model based on anchor frame and anchorless frame,respectively.The former model first preprocesses the images of the tobacco production process using the color information of the images,and then puts the processed images into a two-stage target detection algorithm with an attention mechanism for impurity detection;the latter modeling idea is similar to the former one,and puts the processed images using the color information preprocessing module into a single-stage target detection algorithm without an anchor frame with an attention mechanism for impurity detection.Both algorithms use the same color preprocessing module,but the added attention mechanism module and the underlying model are different.Finally,the recognition effect of the two algorithms in the task of tobacco impurity detection is compared and analyzed.The experiments show that the FSLut model performs the best with an average accuracy of 95.07% and a recall of 98.08%,which can effectively identify the common impurities in the tobacco production process.At the same time,the trained model is embedded into the built cross-platform GUI interface framework and deployed to the edge computing device Jetson Xavier NX,forming a more complete machine vision-based solution for tobacco impurity detection,which will positively contribute to the improvement of automatic detection of impurities in the tobacco industrial production process.