A Class-Weighted Optimization Method Based on Convolutional Neural Network and Data Augmentation with Application to Imbalanced Image Recognition

To address the challenges of class imbalance and feature complexity in paddy pest classification, a Convolutional Neural Network (CNN)-based recognition method was developed. This framework synergistically integrated classical architectures (ResNet, VGG) with strategic implementations of data augmentation and transfer learning, substantially improving model generalizability and classification accuracy. During the data preprocessing phase, diversified augmentation techniques including rotation, scaling, and translation were applied to enhance model adaptability to heterogeneous field conditions. Class-weighted adjustment was specifically implemented to mitigate category imbalance, with particular emphasis on improving performance for minority classes. Subsequent integration of ensemble learning strategies further optimized model robustness, achieving notable improvements in both classification precision (98.2% overall accuracy) and system stability. Experimental validation demonstrated exceptional performance: 96.5% and 95.6% accuracy for dominant pest categories (Cnaphalocrocis medinalis [rice leaf roller] and Chilo suppressalis [asiatic rice burer]), respectively. Notably, the model maintained 93.2% recognition accuracy for the low-sample category Stenchaetothrips biformis (grain spreader thrips). Quantitative evaluations revealed mean precision, recall, and F1-score values of 96.5%,98.4%, and 97.3% across the test dataset, conclusively validating the proposed model's efficiency and robustness for practical agricultural applications.