Research On Intelligent Liquor Selection Method Based On Computer Vision

In the distillation process of liquor production,the liquor selection process can classify and store liquor according to its quality,which can effectively improve production efficiency and the quality of final products.The traditional liquor selection method is to manually observe the visual characteristics of liquor foam produced by the impact of liquor and then classifies liquor.Although the method above is simple and easy to implement,it has the disadvantages of having various judgement standards and relying on human experience,which is not conducive to keeping high liquor quality and hinders the automated production.Therefore,the liquor selection process needs to be automated and intelligent.The intelligent liquor selecting method based on computer vision can extract the visual characteristics of liquor foam in the image,and implement intelligent liquor selection by simulating artificial selection process,which has the advantages of fast judgement,high selection accuracy,low cost,easy to implement and maintain.The method has strong practicability,high application value and research value.The main work and research results of this thesis are as follows:1)The mechanism of liquor selection by vision is analyzed and the intelligent liquor selection algorithm is designed by simulating artificial selection experience.The intelligent liquor selection algorithm consists of image pre-processing algorithm,liquor foam classification algorithm,and post-processing algorithm.2)In the image pre-processing algorithm,a fast circle detection algorithm based on anomalous arc segment culling is designed to detect foreground circular areas.The detection accuracy reaches 97.25%,which is 7.73 percentage points higher than the traditional Huff circle detection.Its running speed is 21 times faster than the Hough circle detection algorithm.3)A lightweight convolutional neural network for liquor foam classification is designed.The network is mainly composed of multi-scale block extracting multi-scale features of images,and SE block calculating the weight of each feature channel.On the test set collected in this thesis,the network's classification accuracy reaches 93.18%,and the computation amount is only 216.10 MFLOPs,which are better than similar networks.After post-processing,the classification accuracy can be further impro ved to 97.59%.4)An intelligent liquor selection system based on servers is developed.Experiment shows that the system only takes 12.55ms to run once on the CPU,and can handle a maximum of 19 production lines simultaneously.It takes 6.23ms to run once on the GPU,and can handle up to 61 production lines simultaneously.5)An embedded liquor selection system based on RK3399Pro is designed by simplifying the liquor selection algorithm,which reduces the cost and the difficulty of installation and maintenance.Combining the hardware characteristics of RK3399Pro,a model pruning method based on feature channel weights is used to simplify the convolutional neural network of the server version algorithm and the algorithm is transplanted to the embedded system.Experiment shows that the model pruning method can reduce the computation amount of the network by 30%without reducing the classification accuracy;the embedded liquor selection system can simultaneously handle 8 production lines.