Glass Refraction Distortion Object Detection via Abstract Features.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Glass surfaces cause reflection and refraction, leading to distorted object features.
• This distortion significantly impacts the accuracy of object detection algorithms.
• Existing methods struggle to effectively mitigate these visual interferences.

Purpose of the Study:

• To develop an object detection algorithm robust to glass-induced distortions.
• To improve the accuracy of object detection in scenarios with glass interference.
• To create a computationally efficient model with reduced parameters.

Main Methods:

• Utilized abstract features for object detection via a novel algorithm.
• Introduced skip connections and expansion modules to reduce model parameters.
• Employed binary cross-entropy loss for abstract feature extraction.
• Implemented a loss function to minimize feature distance between object and source domains.

Main Results:

• Achieved a highest average detection accuracy of 92.57% on the GRI dataset.
• Significantly reduced the number of model parameters to only 5.13 million.
• Demonstrated superior performance compared to state-of-the-art methods like Deep Face and VGG Face.
• The proposed GRI dataset is publicly available on GitHub for further research.

Conclusions:

• The proposed abstract feature-based object detection method effectively overcomes glass reflection and refraction distortions.
• The algorithm offers a high accuracy-to-parameter ratio, making it efficient for practical applications.
• This work provides a valuable contribution to robust object detection in challenging visual environments.