Intelligent Pest Recognition Based on Improved Convolutional Neural Networks with Multi-scale Hybrid Attention Module

Pests represent a formidable challenge to agricultural production, significantly affecting both the quantity and quality of agricultural outputs. The timely detection, prevention, and mitigation of pest damage are paramount to ensuring the quality and safety of agricultural products. However, traditional manual methods for pest identification are not only cumbersome and resource-intensive but also prone to subjectivity, leading to sub-optimal identification efficiency and precision. To address the aforementioned issues, this study introduces an intelligent pest recognition system grounded in a convolutional neural network (CNN). This system integrates a multi-scale hybrid attention module with a residual network, specifically tailored for pest classification. Initially, a comprehensive experimental dataset was compiled by acquiring pest samples through three distinct methods. Subsequently, this dataset was utilized to train the improved CNN. By incorporating a multi-scale hybrid attention module and reconstructing the classifier module within the original ResNet18 model, a specialized CNN named PestNet was developed. Experimental results demonstrate that PestNet attains an average recognition accuracy of 93.61%, marking a significant improvement of 2.04% compared to the baseline ResNet18 network. This study comprehensively validates the effectiveness of the PestNet network model in pest recognition tasks through a series of rigorous evaluation metrics. Further ablation experiments confirm the necessity of each modification in the model and their positive impacts on overall performance. This research not only demonstrates the immense potential and broad application prospects of deep learning in pest recognition but also presents a potential intelligent and convenient technological solution for agricultural pest management.