Learning-Based Shadow Recognition and Removal From Monochromatic Natural Images.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Shadow recognition and removal in monochromatic images is challenging due to lack of color cues and complex natural scene conditions.
• Existing methods struggle with ambiguity from near-black objects and complex illumination.

Purpose of the Study:

• To develop a learning-based scheme for accurate shadow recognition and removal in monochromatic natural images.
• To address the challenges posed by missing chromatic information and complex scene characteristics.

Main Methods:

• Utilized shadow-variant and invariant cues from illumination, texture, and derivative characteristics for shadow recognition.
• Employed a decision tree classifier boosted and integrated into a conditional random field for pixel label consistency.
• Applied a Gaussian model for shadow removal in monochromatic natural scenes.

Main Results:

• Demonstrated accurate identification of shadowed areas in monochromatic images.
• Achieved high-quality shadow-free image recovery through the proposed scheme.
• Outperformed existing state-of-the-art methods in qualitative and quantitative evaluations.

Conclusions:

• The proposed learning-based scheme effectively tackles shadow recognition and removal challenges in monochromatic images.
• The method provides accurate shadow identification and precise recovery of shadow-free images.
• The novel database and evaluation confirm the scheme's efficacy compared to prior art.