LDC: A Long-Tail Data Compensation Method for Pest Detection

In the field of pest detection, the quality of datasets plays a pivotal role in determining the effectiveness of training deep convolutional neural network models. To significantly enhance the accuracy of pest detection, optimizing dataset processing techniques is particularly crucial. While current methods such as resampling and data augmentation can enrich the category class information required by the model to a certain extent and enhance its generalization ability. These methods heavily rely on human judgment, thereby easily introducing subjective biases that can affect the final outcomes. Given this, this paper proposes a Long-Tail Data Compensation (LDC) algorithm for pest detection. The LDC algorithm cleverly integrates the training preferences exhibited by the model during the pre-training stage with the class imbalance in datasets. By adjusting the information distribution across various classes, the algorithm can guide the model to focus more on key information during the training process, which can optimize the training direction. Experimental results on the Insects dataset demonstrate that the LDC algorithm not only effectively mitigates the class imbalance problem but also significantly improves the model's mean Average Precision (mAP), proving the effectiveness of the proposed algorithm in this paper.