This paper presents a glove defect classification method that integrates image enhancement techniques with a lightweight model to enhance the efficiency and accuracy of glove defect classification in industrial manufacturing. A dataset comprising images of five types of gloves was collected, totaling 360 sample images, for the training and validation of a deep learning-based glove defect classification model. Image enhancement techniques, including super-pixels, exposure adjustment, blurring, and limited contrast adaptive histogram equalization, increased dataset diversity and size, improving model generalization. Based on the lightweight model MobileNetV2, the model was improved by reducing the number of input image channels through grayscale conversion and optimizing the loss function. Experimental results demonstrate that the improved MobileNetV2 model achieved an average accuracy of 97.85% on both the original and enhanced datasets, effectively mitigated overfitting phenomena, and exhibited a significantly faster training speed compared to the ResNet34 and ResNet50 models.