Training and Testing Texture Similarity Metrics for Structurally Lossless Compression.

Area of Science:

• Computer Vision
• Image Processing
• Data Compression

Background:

• Texture redundancy in images is a key challenge for efficient compression.
• Existing structural texture similarity metrics (STSIMs) require robust training and testing methodologies.
• Natural texture images contain characteristic perturbations that need to be accounted for.

Purpose of the Study:

• To develop a systematic approach for training and testing STSIMs.
• To enable the exploitation of texture redundancy for structurally lossless image compression.
• To create a data-driven method for training the Mahalanobis formulation of STSIM.

Main Methods:

• A set of image distortions mimicking natural texture perturbations was used for training and testing.
• Empirical studies were conducted to establish a perceived similarity scale for original and distorted textures.
• A data-driven approach was employed to train the Mahalanobis formulation of STSIM using annotated texture pairs.

Main Results:

• Training STSIMs led to significant improvements in metric performance.
• The trained STSIM metrics demonstrated competitive performance compared to state-of-the-art convolutional neural network (CNN) based metrics.
• The proposed method offers substantially lower computational cost than CNN-based approaches.

Conclusions:

• A systematic and data-driven approach enhances the performance of STSIMs for image compression.
• Trained STSIMs provide an efficient and effective alternative to complex deep learning models for texture analysis.
• This research contributes to advancing structurally lossless image compression techniques through improved texture similarity evaluation.