Automated Image Colorization Using Deep Learning

Abstract

Restoring colors to grayscale images has always been a complex challenge in computer vision because a single shade of gray could represent multiple possible colors. Early rule-based and manual methods often gave lifeless or unrealistic outputs. With the rise of deep learning, however, major improvements have been made using CNNs, transfer learning, and generative models. In this work, we present an Efficient Net-based model for automatic image colorization, trained on the large-scale Places365 dataset. To enhance the richness of colors, we introduce a color rebalancing technique that gives more weight to rarely used hues. While traditional evaluation metrics like Mean Squared Error (MSE) remain common, they fail to capture how humans actually perceive image quality. Our results show that the model produces vibrant and realistic colorizations, highlighting both the progress deep learning has enabled and the challenges that remain. We also suggest future directions such as perceptual evaluation metrics, real-time deployment, and user-guided systems.