Saliency detection in the compressed domain for adaptive image retargeting

Area of Science:

• Computer Vision
• Image Processing
• Digital Signal Processing

Background:

• Saliency detection is crucial for image processing tasks like region extraction and resizing.
• Current saliency models operate in the uncompressed domain, which is inefficient for internet images stored in compressed formats (e.g., JPEG).

Purpose of the Study:

• To develop a novel saliency detection model that operates directly within the compressed domain (JPEG).
• To design an adaptive image retargeting algorithm leveraging this compressed domain saliency model.

Main Methods:

• Feature extraction (intensity, color, texture) directly from Discrete Cosine Transform (DCT) coefficients within the JPEG bit-stream.
• Saliency value calculation for each DCT block using Hausdorff distance and feature map fusion.
• Adaptive image retargeting employing block-based seam carving and image scaling, guided by a new texture homogeneity definition.

Main Results:

• Accurate saliency information is derived directly from the compressed domain.
• The proposed retargeting algorithm effectively preserves visually important image regions.
• The algorithm efficiently removes less critical image regions, outperforming state-of-the-art methods.

Conclusions:

• The developed compressed domain saliency detection model enables efficient and accurate analysis of visually important image regions.
• The adaptive image retargeting algorithm significantly improves image resizing by preserving salient content and removing redundant areas.
• This approach offers a substantial advancement for image processing applications dealing with compressed image formats.