A Deep Learning project for image colorization using convolutional autoencoders and generative adversarial networks
Image colorization is an emerging topic and a fascinating area of research in recent years. In this paper, we implemented deep learning algorithms to colorize black and white images. Convolutional autoencoder neural network and generative adversarial network (GAN) have been implemented on black and white images. The networks aim to convert the black and white images to their respective color format. The color format used for the implementation of the networks is RGB color space. The performance of the autoencoder model is evaluated using root means squared error, means squared error, means absolute error and colorization accuracy. The data used for the implementation of the networks consists of 7000 images in the RGB color space. The performance of the implemented autoencoder models and generative adversarial network is measured using mean absolute error. The colorization performance of the generative adversarial network is better that that of the autoencoder models.
Dataset Link: https://www.kaggle.com/theblackmamba31/landscape-image-colorization
Three autoencoders models and one ColorGAN model on a landscape image dataset to colorize black-and-white images to color images are implemented. The two color space chosen for colorization is RGB and CIELAB. The AE-LAB autoencoder model performed better than AE-RGB-1 and AE-RGB-2 autoencoder models in terms of RMSE, MSE and MAE with values 0.0753, 0.0058, 0.0477 respectively. The performance of other two autoencoder models is at par. The colorization accuracy of AE-LAB is better than AE-RGB-1 and AE-RGB-2 with an accuracy of 68.57%. The performance of implemented ColorGAN model is best among all the other models implemented. The colorized images presented above attribute to the better performance of ColorGAN model than autoencoder models. The MAE of ColorGAN model is also better than those of the autoencoder models with a value of 0.006477. The results prove that the implemented ColorGAN’s performance to colorize the grayscale images in better than those of the implemented autoencoder models. The use of deep learning to colorize black-and-white images has a huge potential. This can be attributed by the results shown is this paper. The findings provided in this this paper will help to future researchers to develop deep learning models for image coloration. The autoencoder models as well as ColorGAN model implemented in this project will help to colorize any landscape image with a superior accuracy.