Performance Analysis of Texture Similarity Metrics in Hevc Intra Prediction

The visual signal is highly occupied by regions of homogeneous and repetitive patterns known as Textures. Textures have a common property that their similarity highly deviates from point by point comparison, i.e, two textures can look very similar even if they have some shift, rotation and difference in their distribution. In the context of compression, All of the MPEG reference encoders (including HEVC), aim at minimizing the bitrate at a certain distortion level measured in terms of pixel comparison. For textures, this kind of distortion measure does not usually reflect the amount of perceived distortion. For this reason, we investigate the use of state of the art perceptual similarity metrics as a replacement for this measure. In other words, we aim at optimization the bitrate such that we minimize the perceptual distortion rather that the pixels difference. We used two metrics (Local Radius Index and Structure Texture Similarity Metric) in selecting the best intra prediction mode and block partitioning. Experimental results showed that these metrics try always to retain some structural properties of the textures. These metrics also showed a better rate-distortion performance when the distortion is measured via a distance metric based on texture features.