Single Image Reflection Removal Using Convolutional Neural Networks.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Specular reflection significantly degrades image quality when capturing scenes through glass.
• Existing reflection removal methods often rely on multiple images, which have impractical shooting requirements for general users.

Purpose of the Study:

• To develop an effective single image reflection removal technique using deep learning.
• To address the limitations of multi-image methods by enabling reflection removal from a single photograph.

Main Methods:

• A ghosting model was developed to simulate reflection effects.
• Multiple reflection images were synthesized from a single input image.
• An end-to-end convolutional neural network (CNN) with an encoder-decoder architecture was constructed.
• A joint training strategy with internal and external losses was employed for network optimization.

Main Results:

• The proposed CNN method successfully removed reflections from both synthetic and real-world images.
• The method eliminates the need for handcrafted features and specular filters.
• Achieved superior performance with high scores in Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Feature Similarity Index Measure (FSIM).

Conclusions:

• The developed single image reflection removal method using CNNs is effective and practical.
• This approach offers a significant advancement over traditional multi-image techniques.
• The method demonstrates robust performance and high-quality results in reflection removal tasks.