Loading...

Enhancing Weakly Supervised Semantic Segmentation through Patch-Based Refinement

Javid Tajrishi, Narges | 2024

0 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 57230 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Kasaei, Shohreh
  7. Abstract:
  8. Semantic segmentation, a pivotal topic in computer vision, finds extensive applications across various domains. Leveraging deep networks to tackle this challenge necessitates a large dataset with pixel-level annotations, incurring significant time and cost for preparation. Weakly supervised semantic segmentation utilizes less expensive labels, such as image-level annotations, thereby circumventing the need for costly pixel-level annotations. Consequently, weakly supervised learning has garnered considerable attention. In recent years, vision transformers have achieved significant results compared to convolutional neural networks advancing the state of the art. However, employing vision transformers for semantic segmentation supervised at the image level presents challenges, notably the limited capture of object details and local features in the extracted activation maps. This study introduces a novel approach that integrates local perspectives alongside global views to enhance the learning of local information. By focusing on strengthening the generation of activation maps, the method aims to extract local information learned by the network more completely. Evaluation results demonstrate the efficacy of the proposed activation map extraction technique in improving localization map quality and segmentation accuracy. Moreover, by incorporating the proposed method into a multi-stage framework based on transformers with multi-class tokens, a notable enhancement in segmentation accuracy is observed by minimal computational overhead. This improvement, validated on the PASCAL VOC dataset using the mIoU evaluation metric, demonstrates a significant 1.7% enhancement
  9. Keywords:
  10. Weakly Supervised Semantic Segmentation ; Images Classification ; Deep Learning ; Vision Transformer ; Semantic Segmentation

 Digital Object List

 Bookmark