Soybean Quality Inspection Method Based On Machine Vision

Our country is one of the main soybean production and consumption countries in the world.Soybean grading affects the price of soybeans and the quality of their products,and the precise classification of soybean seeds is an important basis for soybean grading.The traditional two-class classification of soybean seeds cannot meet requirement for fine classification of soybean.In recent years,machine vision has received extensive attention and applications in modern agricultural production and processing because of its non-destructive,highprecision and high-speed characteristics,especially in the detection and classification of agricultural products.Therefore,in response to the application requirements of rapid and multi-classification of soybean seeds,this thesis has developed an online soybean seeds classification system based on machine vision.The system includes three parts:image acquisition,image preprocessing and classification.The main research contents and works are as follows:(1)A set of image acquisition and image preprocessing system is designed.Under the light field illumination,the industrial camera is used to obtain the soybean images.After a series of image processing operations such as median filtering,binarization,watershed algorithm,image and operation,minimum external matrix and image uniform size,the soybean seed images of 227 × 227 pixels is formed,which provides data set for the subsequent classification and recognition of soybean seed images.(2)Aiming at the problems of AlexNet used in soybean seed fast,large number of parameters in multi-classification,large convolution kernel is not conducive to extracting small features of soybean seed image,the parameters of local response normalization are not learning,and Re LU is easy to inactivate.The improvement of AlexNet includes removing the convolution layer and the full connection layer of AlexNet,and compressing the number of neurons in the full connection layer.Using the 3 × 3 convolution kernel as the convolution kernel of the improved AlexNet.The local response normalization is changed to batch normalization.Leaky Re LU is used as the activation function of the improved AlexNet.The improved AlexNet includes 4convolution layers,2 batch normalization layers,3 maximum pooling layers,1 full connection layer and 1 softmax classifier.The improved AlexNet is implemented by multi-threading concurrency,and the improved AlexNet is trained and validated.Its weight is retained as CUDA C++ to realize the weight of the improved AlexNet and test improved AlexNet.After 119 times of training on the server,the classification accuracy of the training set and the validation set of the improved AlexNet is 96.8 %and 91.56 %,respectively.(3)Using CUDA C + + to realize the improved AlexNet,a machine vision system based on NVIDIA Jetson TX2 is built and improved AlexNet is transplanted to this system to complete the online classification system of soybean seeds.The recognition time of a soybean seed on NVIDIA Jetson TX2 is about 6 ms,and the recognition accuracy is 92.92 %,which basically meets the application requirements of rapid classification.