Image semantic segmentation, a task to classify each pixel among the interested classes, is an important problem with a wide range of applications such as autonomous driving, medical diagnosis, industrial automation, and aerial imaging. In recent year...
Image semantic segmentation, a task to classify each pixel among the interested classes, is an important problem with a wide range of applications such as autonomous driving, medical diagnosis, industrial automation, and aerial imaging. In recent years, deep convolutional neural networks have shown outstanding performances in image semantic segmentation. A main bottleneck of these approaches is that it requires large amount of fully-annotated data for training such networks. Since the acquisition of fully-annotated dataset is laborious and expensive, weakly supervised semantic segmentation (WSSS) has been suggested as an promising approach for future research direction. There are various types of weak labels for semantic segmentation, for instance, image-level labels, points, scribbles, and bounding boxes. Among these weak labels, image-level labels are popularly used in WSSS for its simplicity. In essence, image-level label denotes the existence of objects in an image. In this dissertation, we consider the problem of weakly supervised semantic segmentation using image-level label.
In the first part of dissertation, we introduce a new training strategy for weakly supervised semantic segmentation. In the proposed approach, we apply image masking technique inspired by human visual system that focuses on interesting vision field and ignores irrelevant parts. By guiding the attention of classification network using the outputs of the segmentation network, the classification network evaluates the qualities of segmentation output and encourages the segmentation network to generate more accurate output. To boost the segmentation performance, we also introduce simple yet effective technique to train the classification and refine the saliency map. Our experiment results demonstrate that our approach is effective in solving weakly supervised semantic segmentation.
In the second part of dissertation, we introduce a superpixel discovery method that generates semantic-aware superpixels. Our superpixels have new properties that the apart pixels can be grouped into a superpixel if they have similar semantic features. Also, the number of superpixels depends on the complexity of images, not the pre-defined number. Our superpixel expresses semantically similar group of pixels with a very small number of superpixels. We train the segmentation network using superpixel-guided seeded region growing technique which improves the qualities of initial seed. Our extensive experiments show that our approach achieves competitive segmentation performance with the state-of-the-arts in weakly supervised semantic segmentation.