Deep transfer learning based image colorization using VGG19 and CLAHE.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Image colorization is a complex computer vision task due to the ambiguity in mapping grayscale intensity to color.
• Traditional methods often require manual input or reference images, limiting their autonomy.
• Deep learning has emerged as a powerful tool for more reliable and automated image colorization.

Purpose of the Study:

• To present a deep transfer learning framework for high-quality and efficient colorization of grayscale images.
• To leverage advanced deep learning techniques for autonomous image recoloring.
• To improve the visual fidelity and practical applications of image colorization.

Main Methods:

• Utilized a pre-trained Visual Geometry Group (VGG19) network as a backbone for feature extraction, encompassing both texture and semantic information.
• Employed a network architecture with 16 convolutional and 3 fully connected layers.
• Integrated Contrast Limited Adaptive Histogram Equalization (CLAHE) as a post-processing step to enhance image contrast and color vibrancy.

Main Results:

• The proposed framework achieved superior performance, validated by high Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) scores.
• Experimental results on datasets like ImageNet, COCO-Stuff, and Places365 demonstrated outstanding quantitative performance.
• Visual assessments confirmed improved color vibrancy and contrast adjustment compared to existing methods.

Conclusions:

• The deep transfer learning approach offers an effective solution for realistic and efficient image colorization.
• The integration of VGG19 and CLAHE significantly enhances output image quality.
• The method has practical applications in restoring old photographs, enhancing black and white films, and improving medical image visualization.