GAN-Based Image Colorization for Self-Supervised Visual Feature Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Deep neural networks in computer vision require large labeled datasets.
• Manual annotation is costly and time-consuming.
• Self-supervised learning offers an alternative for automatic feature learning.

Purpose of the Study:

• To explore image colorization using generative adversarial networks (GANs) as a self-supervised learning method.
• To leverage GAN-based colorization for visual understanding via transfer learning.
• To improve downstream tasks like multilabel image classification and semantic segmentation.

Main Methods:

• Utilized conditional GANs (cGANs) for realistic image colorization.
• Employed transfer learning to apply colorization features to other tasks.
• Evaluated performance on COCO and Pascal datasets.

Main Results:

• Demonstrated GANs' effectiveness for self-supervised feature learning through colorization.
• Achieved a 5% performance increase in multilabel image classification.
• Showed a 2.5% performance improvement in semantic segmentation.

Conclusions:

• Image colorization with cGANs serves as a viable proxy task for self-supervised feature learning.
• This method effectively boosts downstream computer vision task performance.
• Eliminates the need for manual data annotation in feature learning.