BROWSE
Detail
Conference Paper
MS-STDC: A Multi-Scale Short-Term Dense Concatenate Network for Semantic Segmentation
Cited - time in 
Share 
Share
- Authors
- Yujeong Shin, Mooseop Kim, Chi Yoon Jeong
- Issue Date
- 2023-12
- Citation
- International Symposium on Advanced Intelligent Systems (ISIS) 2023, pp.236-240
- Language
- English
- Type
- Conference Paper
- Abstract
- Semantic segmentation is the task of classifying each pixel in an image into a specific category. To achieve a high accuracy in semantic segmentation, it is necessary to analyze both the contextual and spatial information of images. Methods that utilize multiple branches for extracting specific pieces of information can improve the overall accuracy; however, they are associated with additional computational costs. Therefore, in this study, we propose a multi-scale short-term dense concatenation network (MS-STDC) that leverages the multi-scale information in a segmentation model with a single branch. This network first extracts multi-scale information using a multi-scale backbone module and then fuses the extracted information for each input resolution on a multi-scale basis. For evaluation, the performance of the proposed MS-STDC method was compared with that of a state-of-the-art (SOTA) method on a public dataset. The effects of the multiscale backbone network were validated at three different resolutions: 50, 75, and 100%, of the original image. The experimental results showed that the MSSTDC network outperformed the SOTA model, which had a similar number of layers, regardless of the input resolution. Similar results were obtained when the performance of the network was compared with that of the deeper SOTA model, except when 50% of the resolution of the original image had been used. In particular, the performance of the proposed method improved significantly as the input resolution increased.
- KSP Keywords
- Backbone Network, High accuracy, Multi-scale information, Number of layers, Overall accuracy, Public Datasets, Semantic segmentation, computational cost, short-term, spatial information, state-of-The-Art