Portable Kiwi Variety Classification Equipment Based on Transfer Learning

China is a one of the largest countries producing kiwifruit, and there are many kinds of kiwifruit in China. The identification of kiwifruit varieties has important application value in kiwifruit automatic picking robots and kiwifruit sorting factories. The paper studies the classification of kiwifruit varieties based on deep learning, and studies the deployment and application of kiwifruit classification model on the Jetson Nano artificial intelligence development board. A portable device used for classification of kiwifruit varieties was designed and developed. This paper studied the selection of various parts of the portable device of kiwifruit classification and designed the portable device shell. The portable device shell was obtained through the 3D printer, assembled the parts, and obtained the portable kiwi variety classification equipment. The construction of the kiwifruit dataset of "Cuixiang", "Qinmei", "Xuxiang" and "Hayward" kiwis was studied and using data augmentation to introduce variations into the data. According to the characteristics of the problem of 4 types of kiwifruit classification, the methods and advantages of transfer learning are discussed. Using transfer learning based on three pre-training models Xception, ResNet50, DenseNet121. Comparative analysis of the model size, training speed, convergence, and recognition accuracy of the transfer learning model based on Xception, ResNet50, and DenseNet121, and concluded that in Xception, ResNet50, and DenseNet121, the transfer learning model based on DenseNet121 pre-trained model has the best effect on the classification of four kiwifruit varieties, with fast convergence speed, smallest model, and high recognition accuracy of 97.79%. This paper studied the deployment and experiment of the deep learning model on the Jetson Nano development board. Using the PyTorch to perform transfer learning based on the DenseNet121 pre-training model. After using the TensorRT acceleration engine, it took an average of 0.035s to classify each image, and the real-time classification FPS reached an average of 30, the accuracy rate of the model on the validation set is 93.63%, achieving fast and accurate classification of kiwi varieties.