Data enhancement and multi-feature learning model for pest classification

Abstract

Plant diseases and pests are primary factors that can negatively affect crop yield, quality, and profitability. Therefore, the accurate and automatic identification of pests is crucial for the agricultural industry. However, traditional methods of pest classification are limited, as they face difficulties in identifying pests with subtle differences and dealing with sample imbalances. To address these issues, we propose a pest classification model based on data enhancement and multi-feature learning. The model utilizes Mobile Inverted Residual Bottleneck Convolutional Block (MBConv) modules for multi-feature learning, enabling it to learn diverse and rich features of pests. To improve the model’s ability to capture fine-grained details and address sample imbalances, data enhancement techniques such as random mixing of pictures and mixing after region clipping are used to augment the training data. Our model demonstrated excellent performance not only on the large-scale pest classification IP102 dataset but also on smaller pest datasets.