A pest image recognition method for long-tail distribution problem

Abstract

Deep learning has revolutionized numerous fields, notably image classification. However, conventional methods in agricultural pest recognition struggle with the long-tail distribution of pest image data, characterized by limited samples in rare pest categories, thereby impeding overall model performance. This study proposes two state-of-the-art techniques: Instance-based Data Augmentation (IDA) and Constraint-based Feature Tuning (CFT). IDA collaboratively applies resampling and mixup methods to notably enhance feature extraction for rare class images. This approach addresses the long-tail distribution challenge through resampling, ensuring adequate representation for scarce categories. Additionally, by introducing data augmentation, we further refined the recognition of tail-end categories without compromising performance on common samples. CFT, a refinement built upon pre-trained models using IDA, facilitated the precise classification of image features through fine-tuning. Our experimental findings validate that our proposed method outperformed previous approaches on the CIFAR-10-LT, CIFAR-100-LT, and IP102 datasets, demonstrating its effectiveness. Using IDA and CFT to optimize the ViT model, we observed significant improvements over the baseline, with accuracy rates reaching 98.21%, 88.62%, and 64.26%, representing increases of 0.74%, 3.55%, and 5.73% respectively. Our evaluation of the CIFAR-10-LT and CIFAR-100-LT datasets also demonstrated state-of-the-art performance.