Deep Cleaner—A Few Shot Image Dataset Cleaner Using Supervised Contrastive Learning

Images are increasingly used for AI-based diagnosis and analysis of many diseases like cervical cancer, mouth glucose from retina etc. In cases, data collection is done by specialised camera modules which capture images affected areas. As with any other sources data, this process also error-prone may contain unwanted objects regions that require cleaning removing them. Outliers in these kinds dataset adversely affect the performance machine learning models. Manually would be a tedious task, especially when collated different sources. Hence, before training model utmost importance. paper, we propose Few-Shot based pre-trained supervised contrastive settings to automate cleaning. Our learns distribution distinguishes accurate points noisy points. We show scaling up can greatly improve performance. On MobileODT was collected Kaggle, our obtained 52% accuracy without using an EfficientNet architecture classification task. Whereas same ROI cropping achieved 76.56% after through proposed Deep Cleaner approach requires only 100 clean images. The performs 2.74% better than denoising auto-encoder, considered powerful anomaly detection technique.