Self Supervised Learning Visual Encoder with Mask Aware Gating

Abstract

In the context of self-supervised learning, traditional visual encoders are often limited in processing complex scenes due to the lack of effective feature selection mechanisms. This article proposes a self-supervised learning visual encoder model that combines a mask-aware gating mechanism to enhance the feature learning ability in visual tasks. Specifically, we simulated data loss or noise interference by partially masking input images and designed a hyper self-aware gating module (HSAGM) that adaptively adjusts the model's attention to different features, improving the efficiency and accuracy of feature extraction. In addition, we proposed a mask-based interpolation method that utilizes contextual information learned by the model to make reasonable interpolation predictions for occluded areas. These methods perform well in visual neural coding tasks, improving the training performance of self-supervised learning models. Through experimental verification on benchmark datasets, our method significantly improves the performance of visual tasks such as image classification, demonstrating its potential for application in complex visual scenes.