Pre-Training Study on Image Segmentation based on Multi-Scale Contrast Learning

Abstract

the current popular contrast self-supervised pre-training method is mainly more suitable for the downstream task of global image classification, and for image segmentation of spatial information demanding intensive prediction task, the effect is not satisfactory, so we proposed a multi-scale contrast learning based on image segmentation model, the method in the global, local and pixel three contrast learning, fill the gap between self-supervised pre-training and intensive prediction task. The method for the problem of limited annotation data, efficient use of a small amount of annotation data, data through color enhancement, and then use multiscale contrast loss function training can accurately extract the image features of training model, the training model can in image segmentation this task, according to different specific tasks, fine-tune the overall segmentation network. We demonstrate the effectiveness of the proposed training strategy using the Cityscapes and PASCAL VOC 2012 segmentation datasets. Our results show that pre-training with the proposed contrast loss can achieve high performance gain for the intensive prediction task of image segmentation with limited amount of labeled data, outperforming existing technical methods.