Research On Image Classification Algorithm Based On Contrastive Learnin

Image classification is a fundamental problem in computer vision research,and has a wide range of applications in many fields such as face recognition,medical imaging,traffic control,etc.It is also the basis for other vision tasks such as image detection,object tracking,etc.It is significant to be able to classify images efficiently and accurately.Deep learning methods are capable of extracting image features autonomously and classifying images end-to-end,and have been widely used in image classification tasks.However,traditional supervised learning methods require labeling of the data and require high cost for labeling the data.Based on this,unsupervised contrastive learning has started to be widely studied and applied to image classification problems.This paper focuses on contrast learning based image classification algorithms.The main research components are:(1)The traditional contrastive learning method of data augmentation at the image level to construct the contrastive view is abandoned,and the method of perturbing the encoder parameters and the image representation at the encoder and image representation levels,respectively,is used to construct the contrastive view.When the encoder is perturbed,the encoder parameters are corrected by introducing noise that satisfies the normal distribution.The image representation is perturbed by adding an image history representation,discarding the image representation with a certain probability,and adding representing noise to the image representation,respectively.(2)Clustering is introduced to perform clustering operations on the images to cluster them into different clusters.To solve the widespread problem of false negative samples in contrastive learning,negative samples are sampled only in clusters different from the current one when negative sampling of images is performed.To distinguish the effect of different negative samples on positive samples,the contrastive loss between positive and negative samples is weighted according to the distance between the cluster center of the cluster in which the negative samples are located and the cluster center of the cluster in which the positive samples are located.(3)Combining contrastive learning and supervised learning,the focus on between-samples in contrastive learning is used to compensate for the lack of focus on class-level information but not on information differences between samples in supervised learning.In this way,the cohesiveness between similar images and the separability between different classes of images are strengthened,and the classification boundaries of each class of images in the representation space are made clearer.