基于候选区域生成的弱监督图像语义分割算法

Weakly-supervised Image Semantic Segmentation Algorithm Based on Candidate Regions

针对现有弱监督语义分割方法依赖于初始响应以及分类任务,只关注区分目标对象区域,无法通过完整的区域来优化损失函数的问题。论文提出了一种基于候选区域生成的语义分割算法,用于图像级标注的弱监督图像语义分割。该算法首先在分类网络中引入混合数据增强方案,再通过制定相应的策略,对图像特征进行聚类,构建子类目标并生成子类标签,从而使得训练过程不仅仅依赖于可区分的对象区域。在PASCAL VOC 2012数据集上进行了全面的实验和分析,与其他弱监督语义分割算法相比,论文提出的算法表现出良好的性能:通过使用混合数据增强以及自监督的候选区域生成的方法,使原始图像产生更加完整的响应映射,将交并比(IoU)提高了2.1%,提高了最终分割网络的性能。

The existing weakly-supervised semantic segmentation methods rely on initial response and classification task,only focus on distinguishing target object area,and cannot optimize loss function through complete area.This paper presents a seman-tic segmentation algorithm based on candidate regions for weakly-supervised semantic segmentation of image-level annotation data.In this algorithm,mixed data enhancement scheme is first introduced in the classification network,then the corresponding strategy is formulated to cluster image features,subclass targets are constructed and subclass tags are generated,so that the training process is not only dependent on distinguishable object areas.Comprehensive experiments and analyses are carried out on the PASCAL VOC 2012 dataset,the algorithm shows good performance compared to other weakly supervised semantic segmentation algorithms.By using the method of mixed data enhancement and self-supervised candidate regions generation,the original image produces a more complete response map,which improves the Intersection over Union(IOU)by 2.1%and improves the performance of the final segmenting network.