Learning Content-Weighted Deep Image Compression.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Traditional lossy image compression methods struggle with spatial content variation and contextual dependencies.
• Existing deep learning approaches often face computational challenges and limited parallel decoding capabilities.

Purpose of the Study:

• To develop an efficient and effective learning-based lossy image compression method.
• To address the limitations of traditional entropy models and computationally expensive deep context models.

Main Methods:

• A content-weighted encoder-decoder model with channel-wise multi-valued quantization.
• An importance map subnet for spatially varying code pruning, providing a bitstream length upper bound.
• An upper-triangular masked convolutional network (triuMCN) for efficient large context modeling.

Main Results:

• The proposed method achieves visually superior results compared to existing approaches.
• Favorable performance against both deep and traditional lossy image compression techniques.
• Efficient compression of quantized representations using arithmetic coding.

Conclusions:

• The content-weighted approach effectively balances rate-distortion performance and visual quality.
• The triuMCN enables efficient context modeling for improved compression.
• This method offers a promising direction for advanced lossy image compression.