Unbalanced Classification for Pests Detection and Recognition

To address the problems of long-tailed distribution, scale diversity, and heavy occlusion in the pest identification task under insect detection light, a new method for pest identification with an improved YOLOv5s network, ECA-YOLOv5s, is proposed. First, image weighting and data enhancement strategies are used to address the long-tailed distribution, which increases the effective training samples and greatly alleviates the data imbalance problem while reducing the manual labeling cost. Second, an efficient channel attention mechanism is incorporated into the backbone network to mitigate the impact of dimensionality reduction on the channel attention learning effect. Next, the SPPFCSPC module is introduced in spatial pyramid pooling to extract features from the spatial information of the context. Finally, the decoupled detection head is introduced to separate the classification and regression tasks to improve pest identification accuracy and enhance model robustness. The experimental results show that compared with the YOLOv5s model, ECA-YOLOv5s improves accuracy, recall, and average precision mean values by 15.9%, 17.8%, and 19.7%, respectively. Meanwhile, compared with the YOLOv5m model, ECA-YOLOv5s improves recall and average precision mean values by 17.0% and 14.4%, respectively, while reducing complexity, ultimately reaching 89.1%, 85.8%, and 92.3%. The experiment proves that this method has high recognition and localization accuracy and is suitable for practical deployment in agricultural scenarios.