Dynamic Neural Network for Lossy-to-Lossless Image Coding.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Lifting-based wavelet transform is crucial for visual data compression.
• Coding scheme performance relies heavily on prediction and update lifting operators.
• Current schemes often use linear filters, limiting adaptability.

Purpose of the Study:

• To develop a novel image compression method using neural networks to learn lifting operators.
• To enhance the adaptive capabilities of lifting schemes for diverse image content.
• To improve both lossy and lossless image compression efficiency.

Main Methods:

• Employed Fully Connected Neural Networks (FCNNs) for prediction and update filters.
• Developed a dynamical FCNN model for adaptive learning.
• Proposed two adaptive learning techniques: iterative computation and direct parameter learning via loss function reformulation.

Main Results:

• Demonstrated significant benefits of the proposed FCNN-based lifting schemes.
• Achieved improved performance in both lossy and lossless image compression contexts.
• Validated the effectiveness of adaptive learning techniques on various test images.

Conclusions:

• Neural network-based learning of lifting operators offers superior performance in image compression.
• Dynamical FCNN models provide adaptive compression tailored to image content.
• The proposed adaptive learning strategies effectively enhance lifting-based wavelet transform coding.