TSRRNet: two-stage reflection removal network with reflective guidance.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Reflection removal is crucial in image processing, but training data scarcity necessitates synthesized samples.
• Existing Convolutional Neural Network (CNN)-based models often underutilize reflection information, limiting performance.

Purpose of the Study:

• To design a novel network that refines the transmission layer by effectively utilizing the reflection layer.
• To address the limitations of current CNN-based reflection removal methods.

Main Methods:

• A two-stage generative adversarial network (GAN) architecture is proposed.
• The first stage estimates coarse transmission and reflection layers.
• The second stage refines these layers using gated convolution and a soft mask approach, trained with Wasserstein GAN (WGAN).

Main Results:

• The proposed method demonstrates superior performance compared to several state-of-the-art networks on benchmark datasets.
• Experimental results validate the effectiveness of using reflection information for transmission layer refinement.

Conclusions:

• The novel two-stage GAN effectively refines transmission layers by leveraging reflection information.
• The gated convolution and soft mask approach in the second stage significantly improve reflection removal accuracy.