Research On The Algorithm Of Betel Nut Phase Grading With Fine Grain Size And Characteristic Parameters

The quality of the betel nut is the main factor determining the grade of the betel nut.The classification of the betel nut after the brine process is an important part of the betel nut production process.Different grades have different prices.At present,this link is mainly completed manually.Manual operations are easily affected by workers’ work experience and psychological factors,and it is difficult to be objective and stable,which in turn affects product quality and also brings high labor costs.Factors affecting the appearance of betel nut include shape,such as length,width,etc.,but also texture,such as surface groove shape,groove depth,etc.,and there are also some local features,such as bulge,etc.,some of which are fine-grained,and some are are characteristic parameters.Therefore,this paper studies the classification method of betel nut from two aspects of fine granularity and characteristic parameters,aiming to find a method that can efficiently classify betel nut.The research contents are as follows:(1)Fine-grained recognition algorithm development.This paper draws on the idea of bilinear neural network and proposes a lightweight bilinear attention neural network.The lightweight is reflected in the combination of the first half of the dual-branch network,and the shallow features are extracted through feature map sharing,which reduces the amount of parameters,and then expands into dual branches,so as to extract global features and local features respectively.A spatial attention mechanism is added to improve the feature representation of sensitive parts.In addition,due to the small size of the dataset and the commonality of shallow features,using the transfer learning method to train the network,the final prediction accuracy is 0.88,and the prediction time of a single betel nut is 4.59 ms.(2)Development of Recognition Algorithm Based on Feature Parameters.In this paper,the algorithm is designed by using 8 shape characteristic parameters and 4texture characteristic parameters of betel nut.In order to obtain accurate feature parameters,a rotation method for attitude adjustment of betel nut is proposed in the preprocessing process.In addition,principal component analysis method is used to process the extracted combined features.Finally,a classification model for processing feature vectors is proposed based on deep learning.,In order to determine the model structure,comparative experiments are performed on the model input,whether to add batch normalization,and selecting an appropriate activation function.After the network training is completed,the prediction accuracy is 0.945,and the prediction time of a single betel nut is 14.01 ms.(3)Algorithm recognition effect comparison experiment.Use Res Net and the current mainstream classifier for feature vector processing-support vector machine to do a comparative experiment.The four methods of the lightweight bilinear attention neural network based on fine-grained analysis,the self-built model based on feature parameter analysis,and the Res Net and support vector machine used in the comparative experiments were counted separately.Betel nut predicts time-wise performance.The results show that the self-built model based on feature parameter analysis has better performance when the accuracy is given priority.(4)Algorithm deployment and system development.Based on Lib Torch and C++for algorithm deployment and system development,this paper designs a betel nut recognition system that can be processed in real time at the back end and displayed at the front end.Among them,the back-end algorithm processing module uses a self-built model method based on characteristic parameters.In order to test the real performance of the method,the prediction time,accuracy rate,and CPU usage of a single betel nut when deployed on the machine are counted.