DiWTBR: Dilated Wavelet Transformer for Efficient Megapixel Bokeh Rendering.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Bokeh rendering is traditionally achieved using large-aperture cameras.
• Rendering bokeh from small-aperture images offers system simplicity but poses challenges.
• Existing methods using Convolutional Neural Networks struggle with foreground blurring due to limited receptive fields.

Purpose of the Study:

• To develop an efficient and effective bokeh rendering method for single small-aperture images.
• To overcome the limitations of Convolutional Neural Networks in capturing long-range dependencies for bokeh effects.
• To propose a Transformer-based model that balances performance and computational cost.

Main Methods:

• Proposed a Dilated Wavelet Transformer model for Bokeh Rendering (DiWTBR).
• Employed window attention and dilated attention schemes for local and global spatial interactions.
• Integrated wavelet transform within the attention block to enhance computational efficiency.

Main Results:

• DiWTBR outperforms state-of-the-art methods, achieving up to 0.7dB improvement in Peak Signal-to-Noise Ratio (PSNR).
• The model demonstrates high efficiency, suitable for high-resolution images.
• Requires only 4GB of GPU memory, enabling implementation on standard personal computers.

Conclusions:

• The Dilated Wavelet Transformer (DiWTBR) is a highly effective model for bokeh rendering from single small-aperture images.
• DiWTBR offers a superior balance between performance and computational efficiency compared to existing methods.
• The model's low resource requirements make advanced bokeh rendering accessible on common hardware.