Deep tone mapping network in HSV color space

Deep Tone-Mapping Operator Using Image Quality Assessment Inspired Semi-Supervised Learning

“…Zhang et al [34]'s method converts HDR image into HSV color space. S and V channels are processed by CNN while H is preserved to avoid hue shift during tone-mapping.…”

“…Later work Marnerides et al [39] first applied multi-branch structure by assigning different kernel size in each branch to focus on global/local features. Later, decomposing and multi-branch had become more popular, as they were used by [31], [25], [26], [27], [34], [40], [41], [42], [45], [5] and [47]. Moreover, Wang et al [42] first tackled denoising, while Xu et al [46] first introduced 3D convolution considering temporal information of HDR videos.…”

“…Second, filtering i.e., decomposing image into base and detail components using filter (usually edge-preserving filter like bilateral filter) and respectively sending them into global and local/detail branches ( [31], [42], [45], [5] and [47]). Third, "decomposing" by color channel ( [31], [34] and [40]). We considered the 2 nd idea feasible because task-specific branches can focus on different image components using customized structure.…”

“…This was adopted by almost all deep TMOs due to the wide accessibility of HDR content. However, we found that even when target SDR images are from parameter-fine-tuned TMO (in [25] and [34]) or selected according to best objective score (in [31], [24], [26] and [27]), they still contains artifacts brought by traditional TMOs e.g., over-enhancement or overstylization. Therefore, to avoid those artifacts, we first turned to use simultaneously-shot real HDR-SDR pairs.…”

“…Since there is a huge difference between pixel value distribution of linear HDR and non-linear SDR images, majority of related works followed the common practice of converting HDR images into logarithm (log) domain ( [3], [33], [25], [26], [34], [36], [42], [43] and [46]) or other non-linear domain ( [38]), which made pixel value more evenlydistributed and easier for CNN to process. While the rest of methods ( [31], [24], [27], [39], [40], [41], [45], [5], [47] and [56]) just normalized HDR images without domain transfer.…”