Self-supervised Image Classification Using Convolutional Neural Network

Abstract

Image classification has been a trendy research topic in the field of pattern recognition and computer vision, which extracts different features of images and predict the category of images. Thanks to the development of deep learning, powerful convolutional neural networks can be used in the field of image recognition. However, existing deep learning-based image recognition research mostly follows the framework of supervised learning, and model learning relies on a large number of accurate labels. Providing a large amount of label data will undoubtedly require laborious human effort and expensive costs. Therefore, image recognition based on unsupervised learning (without using any image category labels to achieve classification) has become a spotlight for research. In this paper, explorations on the image classification by self-supervised framework SimCLR on image classification successfully clusters a large number of images into an optimum amount categories. Qualitative results have show SimCLR is particularly effective in recognizing the colors of images; both qualitative and quantitative results shows SimCLR is great at identifying simple contours. However, when the colors are similar, and contour lines are complex, SimCLR does not obtain satisficing results. The accuracy on classifying Mnist dataset is 32%.