Traditional cherry tomato grading machines can only classify fruits based on size and cannot detect the surface defects of cherry tomatoes. To solve the problem, a transfer learning method is used to compare different network models and select the neural network, InceptionV3, which is more suitable for the identification of surface defects of cherry tomatoes. The InceptionV3 model is improved by introducing and modifying the width factor α to reduce the number of channels and improve the training speed of the model. Then the number of neurons in the fully connected layer is modified, and finally the attention network structure is inserted. By comparing different attention network structures, a deep learning model based on the improved InceptionV3 neural network, named N—InceptionV3_FD, is proposed. The results show that the N—InceptionV3_FD model achieves an accuracy of 97.06% in cherry tomato surface defect detection, increased by 7.84% over the original InceptionV3 model, with a more stable loss function maintaining around 0.1. This model provides a theoretical basis for the surface defect detection of cherry tomatoes.

针对传统圣女果品质质量分级机只能通过圣女果果实大小进行分级，不能检测圣女果表面缺陷的问题，采用迁移学习方法，对比不同的网络模型，选出更适用于进行圣女果果实表面缺陷识别的神经网络InceptionV3。以InceptionV3为基本模型进行改进。首先，通过引入并修改宽度因子α来压缩通道数量，提高模型训练速度；然后，修改全连接层神经元个数；最后，插入注意力网络结构并对比插入不同的注意力网络结构，提出一种基于改进InceptionV3神经网络的深度学习模型（N—InceptionV3_FD）。结果表明，N—InceptionV3_FD模型在圣女果表面缺陷检测识别中准确率达97.06%，比原InceptionV3模型提高7.84个百分点，且损失函数值更加平稳，稳定在0.1左右，为圣女果的表面缺陷检测提供理论基础。