关键词: 入侵昆虫, 卷积神经网络, 模型改进, 图像识别

Abstract: The existing insect related recognition algorithms have few kinds of recognition, lack of classification and recognition algorithms for a large number of invasive insects, and are difficult to provide stable and efficient technical support for the recognition function of the integrated invasive insect system. In this study, 31 kinds of invasive insect images are collected, processed and divided into data sets. Based on four convolutional neural network models, DenseNet121, MobileNetV3, ResNet101 and Shuffle Net, training, testing, analysis and discussion are carried out. The results show that MobileNetV3 has better comprehensive performance in the background algorithm application of the identification function of the invasive insect integrated recognition system. According to the existing defects and model characteristics of the MobileNetV3 model, the attention mechanism and activation function of the designated bottleneck layer of the MobileNetV3 model are improved. The accuracy of the improved model is 92.8%, and the average recognition time of a single test set image is 0.012 s, which is 0.5% higher and 15.2% shorter than the original MobileNetV3 model, which can well meet the requirements of multi insect recognition and classification.

Key words:  invasive insects, convolutional neural network, model improvement, image recognition