Advancing Image Classification through Self-Supervised Representation: Comparative Benchmarking with Lightweight CNNs, and Traditional Baselines

Abstract

Image-based plant disease classification has become increasingly important in supporting maintainable agriculture by enabling early detection and intervention. However, existing approaches often face two critical limitations: dependency on heavy data augmentation with the issue caused by class imbalance, when working dataset is limited. This proposed work addresses these challenges through a dual-pipeline framework applied to bean and bean leaf disease classification. The influences self-supervised learning using SimCLR with lightweight CNN backbones (ResNet18 and DenseNet121) to learn robust feature representations directly from raw image distributions, minimizing dependency on large-scale labeling and augmentation at the begining. On the other hand, Sobel-Feldman edge detection and contour analysis were applied through classical image processing with custom-made features contribute to machine learning classifiers such as XGBoost to provide benchmarking baselines model. The dataset contains eight classes, with significant imbalance between majority and minority classes. The experimental results demonstrate that the SimCLR-based pipeline achieves strong generalization for underrepresented classes also and outperforming traditional ML baselines in terms of performance like accuracy, macro f1-score, and precision-recall balance. Here we can say, the classical pipeline contributes clarity and lightweight deployment potential. The experimental findings demonstrates that the effectiveness of self-supervised learning in reducing augmentation dependency, reducing imbalance issue, and providing impartial classification over classes. This work not only improves methodological insights but also delivers a practical and scalable solution for bean disease detection, contributing to future agricultural research and real- world farming applications.