Traditional deep learning-based image classification methods rely on large-scale labeled data. In practical scenarios such as medical imaging and agricultural monitoring, obtaining labeled data requires substantial manpower and time, becoming a bottleneck for deployment. Similarly, in the textile industry, the automated inspection of fabric defects and high-precision fiber texture identification often face similar challenges due to the scarcity of highquality annotated data. Self-supervised learning offers a solution. This study evaluates the effectiveness of contrastive learning-based selfsupervised feature representation in image classification, focusing on feature extraction from unlabeled data. An efficient contrastive learning framework was constructed and evaluated on CIFAR-10, ImageNet-1K, and CUB-200-2011. Based on the ResNet architecture, combined with the InfoNCE loss and data augmentation, a two-stage training strategy of self-supervised pre-training followed by supervised fine-tuning was adopted. Model performance was assessed using classification accuracy and F1 score. Results show that the proposed method outperforms traditional supervised learning, especially in low-label regimes and as a robust initialization strategy across datasets. The model demonstrates strong generalization in low-sample settings and adaptability to different data distributions. This study clarifies the role of contrastive learning in feature representation for image classification and provides support for applying self-supervised learning in domains with limited annotations, such as medical image analysis and agricultural monitoring. It also offers a transferable framework for related computer vision tasks.

textile manufacturing, contrastive learning, self-supervised learning, feature representation, image classification