Model Based Image Compression and Adaptive Data Representation by Interacting Filter Banks

NeurIPS 1989

TL;DR

This paper proposes a novel image compression technique that leverages a model of the human visual system, specifically its multi-channel characteristics and inhibitory interactions. By combining a Markov Random Field (MRF) image model with an adaptive data representation based on a vast number of interacting filter banks, the authors achieve high-quality image compression with a promising signal-to-noise ratio.

What Problem Does This Paper Address?

Achieving high-rate image data compression while maintaining high reconstructed image quality is a persistent challenge. Traditional methods often struggle to balance compression ratios with visual fidelity. This paper tackles this by proposing a model-based approach that draws inspiration from the human visual system's processing of visual information. The core problem is to efficiently represent image data in a way that is both compact and allows for faithful reconstruction, especially considering the inherent ill-posed nature of reconstructing an image from compressed data.

Methods

The authors introduce a mathematically coherent parallel architecture for image data compression. Key components include:

  • Markov Random Field (MRF) Image Model: This model is used to address the ill-posedness of image reconstruction by incorporating prior knowledge about image structures. The mean field approximation of the MRF is linked to recurrent neural networks.
  • Adaptive Data Representation: Unlike previous methods that used fixed representations, this paper proposes an adaptive approach. This representation is tuned to each specific image through a process of "competition and cooperation" among a large number of filter banks.
  • Interacting Filter Banks: A vast number of filters are prepared, centered at each site and varying in size, shape, and orientation. Two-dimensional Gaussian filters are used for DC components (luminance), and their first-order derivatives represent gradients. Crucially, filters whose receptive fields significantly intersect with the "line process" (representing image discontinuities) are inhibited. Inhibitory interactions are also introduced between filters of similar characteristics at nearby sites.
  • Line Process: This component represents image discontinuities, such as contours or boundaries between objects, and aids in segmenting the image into regions.
  • Stochastic Relaxation: Both at the encoder (for parameter estimation) and decoder (for reconstruction), stochastic relaxation is employed. This is a form of energy minimization, akin to simulated annealing, to find optimal configurations.

Datasets

The paper mentions using "various images" for initial testing of the energy learning rule. For the experimental results presented, they used an image dataset consisting of 256x256 pixels, with each pixel having 8-bit gray levels. No new benchmark datasets were introduced.

Key Insights and Ideas

  • Biologically Inspired Compression: The core insight is that by modeling the human visual system's multi-channel processing and inhibitory interactions, more effective image compression can be achieved.
  • Adaptive Filtering for Representation: The idea of using a dynamic, adaptive set of filters that compete and cooperate to represent image data is a significant departure from fixed filter banks.
  • Line Process for Discontinuity Handling: Explicitly modeling image discontinuities with a "line process" is crucial for effective segmentation and subsequent processing.
  • MRF for Ill-Posedness: The use of MRFs provides a principled way to handle the inherent ambiguity in image reconstruction.

Key Results

The paper reports that their proposed energy learning rule works well for various images. For a specific experimental example (shown in Fig. 2a), the reconstructed image (Fig. 2b) achieved a signal-to-noise ratio (SNR) of approximately 38dB after 20 iterations of relaxation computation. The figure also illustrates sampled data points and selected filter shapes, highlighting the adaptive nature of the sampling and filtering.

Limitations and Future Work

The authors acknowledge that the regularization parameter is increased to infinity if filter outputs are not further compressed, which might be a limitation for achieving even higher compression ratios. Future work could explore further compression of filter outputs. The paper focuses on a single example for experimental results, suggesting that more extensive experiments across a wider range of images would be beneficial to fully assess the scheme's potential.

Summary updated

BibTeX
@inproceedings{1989model,
  title = {{Model Based Image Compression and Adaptive Data Representation by Interacting Filter Banks}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1989},
}
Figure