DeFusionNET: Defocus Blur Detection via Recurrently Fusing and Refining Discriminative Multi-Scale Deep Features.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Image defocus blur detection faces challenges including background clutter, scale sensitivity, and loss of boundary details.
• Existing methods struggle with complex visual scenes, necessitating improved feature fusion and refinement techniques.

Purpose of the Study:

• To propose DeFusionNet, a novel deep neural network for robust defocus blur detection.
• To address limitations of current methods by enhancing multi-scale feature fusion and refinement.

Main Methods:

• Developed DeFusionNet, a recurrent neural network that fuses and refines multi-scale deep features.
• Implemented feature adaptation and channel attention modules to improve feature discrimination and detail preservation.
• Utilized fused shallow and semantic features, propagated recurrently between network layers for enhanced blur localization and detail refinement.

Main Results:

• DeFusionNet demonstrated superior performance in defocus blur detection across multiple datasets.
• The network effectively handled background clutter and scale variations, improving boundary detail accuracy.
• Experimental results validated the efficacy and efficiency of the proposed DeFusionNet model.

Conclusions:

• DeFusionNet offers a significant advancement in image defocus blur detection.
• The recurrent fusion and refinement of multi-scale features effectively overcome existing challenges.
• The proposed method provides a robust solution for real-world applications requiring precise blur analysis.