Research On Weakly-supervised Semantic Segmentation With Global Discriminative-Region Detection

The recent segmentation algorithms of weakly supervised images usually generate the initial localization map based on the weak localization of the general classification network,and then expand the object semantic area in the initial localization map fora more complete one based on some erasing or area growth method.The segmentation network is trained based on the location map for synthesis of segmentation labels.This idea can obtain good localization areas of the object,but there are still some limitations in the accuracy,training steps and location efficiency.This thesis targets at these shortcomings by exploring image segmentation based on image-level tags.A localization map optimization method based on global region growth,and its improvement with a label quality updating and multi-scale information optimization techniques are proposed to improve the semantic segmentation performance.(1)A localization map optimization method for global region growth is proposed to solve the sparse problem of t the localization map,Here,a global region growth network(GRGN)for segmentation was designed.It processes the feature map globally based on the localization distribution of the initial localization area,highlighting the semantic area where the object is not significant.Then,the significant semantic regions are used to guide the network to focus on this area during training and obtain a more complete object semantic area by activating it and integrating with the initial localization map.Finally,segmentation labels are synthesized based on the localization map to complete the segmentation learning.(2)A segmentation label quality improvement and multi-scale information optimization method is proposed to overcome the inaccurate segmentation label information and insufficient semantic information.This method is integrated with the global region growth network proposed in Chapter 3 with the Spatially Cascaded Pyramid Network(SCPN).First,the quality of segmentation labels is optimized by merging the positioning map dynamically with dynamic thresholding to synthesize the segmentation labels.The dynamically fused positioning map makes the contrast between the foreground and background information in the fused positioning map significant and the edges prominent,while the synthesis of segmentation labels by dynamical foreground-thresholding makes the synthesized segmentation labels during training more complete and accurate than previous methods.Then,the spatially cascaded pyramid module based on the hollow convolution design fulfills the multi-scale optimization.The cascade growth and layered fusion structure in the module can mine more object semantic information and output robust segmentation prediction maps.Finally,images of the segmentation prediction can learn more semantic information under the supervision of segmentation labels.