We propose the Uniform Selection and Representation Matching framework, an image classification framework that leverages co-teaching, contrastive learning, representation matching, and uniform selection to perform accurate wound stage classification with limited and noisy-labeled data. Given that descriptors of wound stages are under-specified, making accurate recognition difficult, images that generate low classification confidence are identified using an entropy-based selection process. Pseudo-labels are assigned to the low-confidence images through the representation matching process, where images are embedded into latent space and labels are assigned through majority voting. The Uniform Selection and Representation Matching framework demonstrates high accuracy in classifying wound-stage images by achieving a classification accuracy of 90.0%, a significant improvement over conventional convolutional neural networks.Clinical relevance- This work proposes a wound-stage classification algorithm trained with minimal data and noisy labels. Applications include remotely monitoring wound healing, recommending treatments, and incorporating intelligent bandage devices.