A Convolutional Neural Network-Based Quantization Method for Block Compressed Sensing of Images.

Area of Science:

• Computer Vision
• Signal Processing
• Machine Learning

Background:

• Block compressed sensing (BCS) is vital for resource-limited image/video coding.
• Quantization of BCS measurements presents challenges, causing errors and redundancy.
• Existing methods struggle with efficient and accurate BCS measurement quantization.

Purpose of the Study:

• To propose a new quantization method for BCS measurements using CNNs.
• To minimize quantization errors and encoding redundancy in BCS.
• To improve the information content of quantized BCS data.

Main Methods:

• Developed a CNN-based quantization and dequantization process for BCS measurements.
• Trained quantization and dequantization CNNs jointly.
• Utilized block distribution parameters as side information, quantized with 1 bit.

Main Results:

• Achieved an average PSNR improvement of 0.48 dB compared to uniform quantization and entropy coding.
• Demonstrated effectiveness across four public datasets.
• Improved coding efficiency at a compression bit rate of 0.1 bpp without entropy coding.

Conclusions:

• The proposed CNN-based quantization method effectively reduces errors in BCS.
• This approach offers superior performance for image/video coding applications.
• The method provides significant PSNR gains, particularly at low bit rates.