[Paper Review] Hybrid Approaches to Image Coding: A Review

Rehna. V. J, Jeyakumar. M. K|arXiv (Cornell University)|Sep 12, 2012

Advanced Data Compression Techniques33 references3 citations

TL;DR

This paper reviews hybrid image coding techniques that combine classical methods like predictive coding, transform coding, and vector quantization with modern approaches such as wavelets and neural networks. It highlights the neuro-wavelet approach—using discrete wavelet transforms followed by neural network-based compression—as a highly effective method that achieves high compression ratios and superior image quality, especially at low bit rates, outperforming traditional standalone methods.

ABSTRACT

Nowadays, the digital world is most focused on storage space and speed. With the growing demand for better bandwidth utilization, efficient image data compression techniques have emerged as an important factor for image data transmission and storage. To date, different approaches to image compression have been developed like the classical predictive coding, popular transform coding and vector quantization. Several second generation coding schemes or the segmentation based schemes are also gaining popularity. Practically efficient compression systems based on hybrid coding which combines the advantages of different traditional methods of image coding have also been developed over the years. In this paper, different hybrid approaches to image compression are discussed. Hybrid coding of images, in this context, deals with combining two or more traditional approaches to enhance the individual methods and achieve better-quality reconstructed images with higher compression ratio. Literature on hybrid techniques of image coding over the past years is also reviewed. An attempt is made to highlight the neuro-wavelet approach for enhancing coding efficiency.

Motivation & Objective

• To analyze and compare various hybrid image compression techniques that integrate classical and advanced methods.
• To identify the limitations of standalone compression techniques in achieving high compression ratios without sacrificing image quality.
• To evaluate the effectiveness of combining wavelet transforms with neural networks (neuro-wavelet approach) for enhanced image coding performance.
• To provide a comprehensive review of literature on hybrid coding systems, emphasizing recent advancements and performance metrics.
• To demonstrate that hybrid systems, particularly neuro-wavelet models, offer superior trade-offs between compression ratio, computational efficiency, and reconstructed image quality.

Proposed method

• The paper reviews hybrid coding systems that combine two or more traditional image compression techniques—such as predictive coding, transform coding, vector quantization (VQ), and segmentation-based methods.
• It evaluates the integration of discrete wavelet transform (DWT) with neural networks, where DWT decomposes the image into subbands, and neural networks compress each subband based on its frequency characteristics.
• For each subband, different neural network architectures (e.g., 8-6-8, 8-4-8, 16-1-16) are applied depending on the frequency content and energy distribution.
• The hidden layer outputs of the neural networks are scalar quantized and then entropy coded using Huffman coding to further reduce bit rate.
• The method uses bit rate (bpp) and PSNR as key performance metrics, with PSNR calculated using the standard formula: PSNR = 20 * log10(255 / sqrt(MSE)).
• The approach applies DPCM to the low-frequency band (band1), while higher-frequency bands (band2–band6) are compressed using specialized neural networks, and band7 is discarded due to low energy.

Experimental results

Research questions

• RQ1 How do hybrid coding techniques that combine multiple traditional methods improve compression ratio and image quality compared to individual methods?
• RQ2 To what extent does integrating wavelet transforms with neural networks enhance image compression efficiency and reconstruction quality?
• RQ3 What is the performance of the neuro-wavelet model in terms of PSNR and bit rate across different wavelet filters and test images?
• RQ4 How does the choice of neural network architecture per subband affect the overall compression performance?
• RQ5 Can hybrid systems reduce computational complexity while maintaining high-quality image reconstruction?

Key findings

• The neuro-wavelet hybrid approach achieved a PSNR of 29.371 dB at a bit rate of 0.40 bpp using the db18 wavelet filter on the Lena image.
• The db6 filter yielded a PSNR of 29.25 dB at 0.39 bpp, and the db4 filter achieved 29.058 dB at 0.37 bpp, showing a trade-off between filter complexity and performance.
• The proposed neuro-wavelet model outperformed both classical wavelet-based compression and standalone neural network compression in terms of image quality and processing speed.
• Wavelet-based hybrid techniques produced high-quality reconstructed images even at low bit rates, particularly effective below 75% compression.
• The integration of neural networks with wavelet transforms significantly improved compression efficiency, especially in low-bit-rate regimes.
• The use of different neural network architectures per subband (e.g., 8-6-8 for band2–3, 16-1-16 for band5–6) enhanced performance by adapting to varying frequency content and energy distribution.