| Title |
An Improved Trans-Unet with Group Attention for Single Image Haze Removal |
| Authors |
홍찬의(Chan Eui Hong) ; 최현덕(Hyun Duck Choi) |
| DOI |
https://doi.org/10.5573/ieie.2022.59.6.46 |
| Keywords |
Computer vision; Image dehazing; Vision transformer; Unet; Group attention block |
| Abstract |
With the recent development of computer vision technology, computer vision technologies such as artificial intelligence-based object detection and image segmentation are attracting attention in the field of autonomous driving. However, these technologies degrade performance due to image loss in driving environments under adverse weather conditions such as nighttime, heavy rain, and fog, which can cause fatal human casualties. In this paper, we propose a novel Group Attention Block (GAB) and a haze removal model combined with the Unet and Vision Transformer in order to get robust computer vision technologies even in such adverse image conditions. This network can capture the image feature with spatial information by CNN layer as well as capture global relations without inductive bias through Vision Transformer. Finally, GAB enhances these functions and helps the decoder to restore the clean image. The simulation results show an improvement in dehazing by PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity Index Measure) score compared to the previous image dehazing models |