Efficient Pest Classification Using Lightweight Neural Networks for Sustainable Pest Control

The existence of pests and diseases threatens the agricultural production and brings serious economic losses to countries whose economy is based on agriculture. Traditional agricultural production methods require a lot of manpower to identify and detect pests and diseases, which will increase the cost of agricultural production. Now, with the development of deep learning technology, computer vision based approaches can help reduce costs and improve efficiency. Through the deep learning model based on CNN, insects and pests can be accurately detected and recognized. This paper proposes an improved recognition model, MFnet, based on the lightweight deep neural network MobileNetV3. A new activation function, MLU6, is introduced in the proposed MFnet model. At the same time, we use the PolyLoss function to replace the cross-entropy loss function when training the proposed model. We also use transfer learning to improve the efficiency of network optimization and prevent overfitting through data augmentation and regularization. Ablation and comparative experiments are carried out and the IP102 pest dataset is used in the experiments. The experimental result shows that MFnet outperforms the original Mobilenetv3 and the other classic models in terms of classification accuracy, at the cost of increased number of parameters compared to the original Mobilenetv3. However, the parameter number, training time and classification time of the proposed model are still much less than those of the other classic models.