关键词: 核桃仁分类, 图像预处理, 卷积神经网络, 推理优化

Abstract: In order to further improve the accuracy and realtime performance of walnut classification recognition, an algorithm model based on the convolutional neural network was proposed, which included an input layer, a hidden layer composed of four convolutional and pooling layers, two fully connected layers, and an output layer. The generalization and robustness of the convolutional network model were enhanced by the optimal design of parameters and shortened by inference optimization with the OpenVINOTM（Open Visual Inference & Neural Network Optimization） tool suite. The trained convolutional network model was inferentially optimized and experimentally validated. The test results showed that the recognition accuracy of the convolutional network model without inference optimization was 98.31%, and the average real recognition time was 323.2 ms. The recognition accuracy of the convolutional network model with inference optimization was 99.44%, and the average real recognition time was 29.6 ms. Compared with the convolutional network model without inference optimization, the accuracy and realtime performance were improved, and the execution time was increased by about ten times. The algorithm model of reasoning optimization can better meet the classification and recognition of walnut kernel and has certain applicability.

Key words: walnut kernel classification, image preprocessing, convolutional neural network, inference optimization