HDR-Like Image Generation to Mitigate Adverse Wound Illumination Using Deep Bi-directional Retinex and Exposure Fusion

Hou, Songlin

Etd

HDR-Like Image Generation to Mitigate Adverse Wound Illumination Using Deep Bi-directional Retinex and Exposure Fusion

公开

Periodic assessment is necessary to monitor the healing progress of chronic wounds. Image analysis using computer vision algorithms has recently emerged as a viable alternative as demonstrated by prior work. However, the performances of such image analysis methods degrade on images captured in adverse illumination that occurs in many indoor environments. To mitigate these lighting problems, High Dynamic Range (HDR) image enhancement techniques can be used to mitigate over- and under-exposure issues and preserve the details of scenes captured in non-ideal illumination. In this paper, we address over- and under-exposure simultaneously using a deep learning-based bi-directional illumination enhancement network based on the Retinex theory. Over- and under-exposed images are generated, which are then fused into a final image with enhanced illumination in an exposure fusion step. Our proposed method outperformed the state-of-the-art on various metrics including structure similarity, peak signal-noise ratio and changes in segmentation accuracy (SSIM scores 0.76/0.69 on bright/dark images, PSNR scores 28.60 on dark images and DSC scores 0.76/0.74 on bright/dark images).

Creator