https://scholars.lib.ntu.edu.tw/handle/123456789/637083
Title: | Deep Learning Based End-to-End Specular Reflection Removal for Medical Endoscopic Images | Authors: | CHI-SHENG SHIH Liao, Yu Cheng CHING-TING TAN |
Keywords: | CNN | Endoscope Image | ResNet | Issue Date: | 6-Aug-2023 | Source: | 2023 Research in Adaptive and Convergent Systems RACS 2023 | Abstract: | Endoscopic images usually have many overexposed regions due to strong and focused light sources and, consequently, physicians need to change the camera angle for a clear view from time to time. This work targets removing the specular reflections from the endoscopy images but maintaining the structural and semantic texture in the images. Unlike the state-of-the-art image inpainting methods, the proposed method treats the specular reflection removal problem as an image-to-image translation task by skipping the specular detection step. The technique uses Dense Multi-scale Fusion Network (DMFN), which aims for small size and clear images, as the base and revises the network to support the large size and noisy endoscopy images. Moreover, the method can recover structure from large overexposed regions. Experimental results show that under different glare ratios and granularity settings, it is more effective than the Partial Convolution Model when recovering the corrupted pixels: the mean average error of pixel values decreases 13.2%, and the PSNR value increases 2.62dB. Qualitative analysis shows that even in the worst case when the proposed model generates output images with MAE up to 4.30, its recoveries still possess good visual qualities. The proposed model can process a 480 × 768 image in 16.2ms on a computer equipped with 36 CPU cores and an RTX-3090 GPU. It usually takes at least 20 minutes for a human to recover such images. |
URI: | https://scholars.lib.ntu.edu.tw/handle/123456789/637083 | ISBN: | 9798400702280 | DOI: | 10.1145/3599957.3606226 |
Appears in Collections: | 醫學系 |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.