Viktoriia Reznichenko, Hyo Young Shin, Kye Dong Jung, Cheol Young Go
언어
영어(ENG)
URL
https://www.earticle.net/Article/A481192
※ 원문제공기관과의 협약기간이 종료되어 열람이 제한될 수 있습니다.
원문정보
초록
영어
We designed a multifunctional neural network to jointly perform semantic segmentation and object boundary extraction. Our goal was to enhance boundary precision, especially for thin and complex facial structures. The model is built on a pre-trained ResNet101 backbone and incorporates ASPP and transposed convolutions to handle objects across multiple scales. A parallel branch structure enables simultaneous learning of semantic regions and boundary details. To further improve visual clarity, we introduce an RoI Tanh-warping technique that selectively distorts the background while preserving the natural appearance of the target region. We also apply a progressive layer unfreezing strategy to allow smooth adaptation to new tasks while retaining key pre-trained features. Experimental results confirm that our method delivers superior boundary accuracy, achieving mean F1 scores of 93.40 and 93.42 on the validation and test datasets. These findings demonstrate that the proposed approach provides both quantitative gains and visually coherent segmentation results.
목차
Abstract 1. Introduction 2. Related Works 3. Theoretical background 3.1 Boundary-Aware Semantic Segmentation (BASS) 3.2 Image Transformation: RoI Tanh-Warping 3.3 Progressive Unfreezing process 3.4 Fine-tuning process 4. Results 5. Conclusions 6. Future Work Acknowledgement References