Low Complexity Pixel Domain Perceptual Image Compression via Adaptive Down-Sampling
Zhe Wang, Sven Simon · 2016
Summary form only given. Compression is becoming imperative for many real-time high-throughput embedded imaging systems. Such systems often have limited computational power and memory resources due to the tight constraints on latency and hardware cost. Transform based codecs like JPEG, JPEG 2000 and JPEG XR require storing multiple image lines, resulting in a non-negligible latency as well as expensive on-chip memories. While the JPEG-LS standard [1] provides a reasonable balance between complexity and compression ratio for lossless coding, its use in near-lossless coding is limited by issues such as stripe-like visual artifacts in smooth image regions. This paper proposes a low complexity image coder that exploits pixel domain perceptual redundancies. The algorithm is based on adaptive down-sampling guided by a pixel domain just-noticeable distortion (JND) model [2] which estimates the threshold below which a difference to the original pixel can be masked by image activities nearby. The input image is down-sampled in both dimensions and the distortion visibility thresholds given by the JND model are determined for all pixel locations in the down-sampling proximity. If the differences between the down-sampled pixel and all original pixels do not exceed the corresponding JND thresholds, the down-sampled pixel is encoded using algorithmic tools similar to those in JPEG-LS. If any of the differences exceeds the JND threshold, all original pixels in the current proximity are encoded with a quantization parameter selected based on the JND model. Such a down-sampling approach can be considered as adaptive to region-of-interest (ROI) identified by comparing the distortions with the JND thresholds. The binary ROI map is compressed with extended Huffman code and transmitted to the decoder. Experiments on 8-bit ITU-T test images on average show up to 20% gain in coding efficiency over JPEG-LS at about 2 bits-per-pixel (bpp) for the same MS-SSIM.