Accurate pest image diagnosis is a critical component of modern precision agriculture. It is important to protect agricultural production, sustainability, and economic progress. Although deep convolutional neural networks, such as ResNet-50, have been widely used for pest classification. However, their performance remains limited when applied to complex, imbalanced, and real-world pest datasets. In this paper, we present an optimized ResNet-50 pipeline for pest image classification on the IP102 dataset, leveraging advanced data preprocessing, targeted regularization strategies, and hyperparameter tuning to improve model robustness and generalization. Moreover, there is substantial bias toward majority classes, and class-imbalance adaptation approaches (e.g., weighted loss functions) have been applied to alleviate this issue. The proposed approach incorporates Mixup data augmentation, label smoothing, exponential moving average (EMA), and stochastic weight averaging (SWA) to enhance training stability and generalization under severe class imbalance. Experimental results demonstrate that the refined ResNet-50 model is a compact and effective CNN that outperforms several related works on the IP102 dataset, achieving 76% accuracy and a weighted F1-score of 0.75. The findings highlight the importance of systematic refinements in enhancing the effectiveness of conventional deep learning models in pest image diagnosis and agricultural tasks.