Semi-supervised Semantic Image Segmentation Based On Bounding Box

With the development of industrial automation and artificial intelligence,the application of semantic image segmentation is widely used in all walks of life.However,how to efficiently and accurately segment different semantic objects in an image is still a technical difficulty in the related industrial field.Recently,Deep Convolutional Neural Networks(DCNN)has made great breakthroughs in the semantic segmentation field,making semantic segmentation possible.At present,there are several major challenges in semantic segmentation,the most prominent are the difficulty to obtain pixel-level labels and the low accuracy of the segmentation model.Therefore,our paper will focus on these two issues.For the difficulty of obtaining pixel-level labeled samples,we solve it by weakening the excessive dependence of model on pixel-level label.Due to pixel-level labels are required for training segmentation model,which is very costly to obtain,while the acquisition of bounding box label is much easier.Therefore,the semi-supervised semantic segmentation with bounding box label has great research significance.In this paper,we explore the segmentation performance of semi-supervised learning when using boundary box labels to supplement pixel-level labels.The experimental results show that when the coarse pixel-level label obtained by the bounding box label is more accurate,the weak sample can completely supplement the loss caused by the reduction of strong sample,and the segmentation performance of semi-supervised model can be improved when the number of weak samples is sufficient.For the low precision of the semi-supervised model,this paper proposes a new method to refine the bounding box label.In the semi-supervised field of bounding boxes,the main challenge is how to convert bounding box labels to coarse pixel-level labels meticulously.To this end,this paper proposes a novel method to refine the bounding box annotations by iterative mining,cleaning,probing and repairing,which can convert the bounding box label into coarse pixel-level more accurately.The coarse pixel-level labels are updated by the more refined labels continuously,and then combined with the pixel-level labels jointly to train the DCNN model.Especially,we use the generated model to mine the information in the bounding box labeled samples,and the mined labels are cleaned and repaired as the refined coarse pixel-level labels The experimental results show that the segmentation performance can be improved by refining the bounding box labels continuously.The experimental results on the PASCAL VOC 2012 standard dataset verify the effectiveness of the proposed method in semi-supervised semantic segmentation,especially for the case of fewer pixel-level labels.