High Frequency Detail Accentuation in CNN Image Restoration

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Machine learning methods, particularly Convolutional Neural Networks (CNNs), often neglect rare but significant events due to their statistical inference nature.
• This oversight leads to unsatisfactory performance in image restoration tasks, resulting in blurred details and loss of critical information.

Purpose of the Study:

• To develop a new training methodology that enhances CNN sensitivity to atypical but semantically important features in image restoration.
• To overcome the inherent limitation of existing CNNs in handling rare events, thereby improving image sharpness and clarity.

Main Methods:

• Proposed a "high frequency feature accentuation space" to promote image sharpness and clarity.
• Incorporated an auxiliary loss term during training to enforce agreement between ground truth and restored images in this feature space.
• Focused on maximizing discrimination of atypical features to penalize image blurs.

Main Results:

• The new CNN design method was implemented and tested on image super-resolution and denoising tasks.
• Experimental results demonstrated significant improvements in image sharpness and clarity.
• The method successfully sensitized CNNs to desired, albeit atypical, features.

Conclusions:

• The proposed training methodology effectively addresses the issue of blurred details in CNN-based image restoration.
• The high frequency feature accentuation space is a viable approach to improve CNN performance on critical, rare events.
• This work offers a promising direction for developing more robust and accurate image restoration techniques.