Performance comparison of image enhancers with and without deep learning

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Image enhancement algorithms aim to improve the visibility of image details and content.
• Traditional methods rely on physical models, while deep learning techniques learn from data.
• Poorly illuminated images present a significant challenge for visual analysis.

Purpose of the Study:

• To empirically compare traditional and deep learning image enhancement methods.
• To evaluate the impact of these enhancement techniques on image retrieval performance.
• To promote informed usage of image enhancers in computer vision applications.

Main Methods:

• Selection of representative traditional and deep learning image enhancement algorithms.
• Experiments conducted on public datasets featuring poorly illuminated images.
• Evaluation of image enhancement techniques using image retrieval algorithms (SIFT, ORB).

Main Results:

• All tested enhancers improved image visibility and detail.
• Deep learning methods generally produced images with lower noise levels compared to traditional methods.
• Noise characteristics of enhanced images significantly affected image retrieval accuracy for SIFT and ORB.

Conclusions:

• Deep learning-based image enhancement offers advantages in noise reduction for poorly illuminated images.
• The choice of enhancement method should consider the noise sensitivity of downstream computer vision tasks.
• Understanding enhancer impact is critical for optimizing applications like image retrieval.