Evaluation of the Encoding Accuracy of the PQ based HDR Content Delivery Formats

Recently, new formats like HDR10 and Dolby Vision have emerged and become the popular choice for delivering HDR content. HDR10 uses a bit-depth of 10 to quantize the data after applying a non-linear transformation operator known as the perceptual quantizer (PQ). It has been known that at this precision of just 10 bits, banding artifacts can be observed in the lower luminance regions. However, in general, HDR10 has good visual quality on HDR displays and is widely recognized by the display makers and content developers. This paper evaluates the accuracy of the content encoded in HDR10 format, especially when the dynamic range is high. It is observed that the accuracy of HDR10 is low when there are extremely dark and bright regions in the image. Therefore while it is the right choice for HDR display, other formats should be used to preserve the original dynamic range for applications such as machine-based analysis and tone-mapping for LDR display. We also propose a dual-layer codec using companding, which resolves the known drawbacks of two-layer backward compatible formats. The proposed format does not require metadata for the reconstruction of the HDR image, and therefore its decoding function has a low computational cost. Moreover, its encoding accuracy is higher than HDR10 and the existing state of the art popular formats like JPEG-XT and HDR-MPEG.