An edge-refined vectorized deep colorization model for grayscale-to-color images

To handle the colorization problem, we propose an edge-refined vectorized deep colorization model. We discuss the reasonable network parameters like the patch size, amount of layers, convolutional kernel size and amount. To improve the colorization performance and simplify the model, two neural networks are respectively trained to obtain the value of U and V components since the model is in YUV color space. In the training stage, we alternately apply two loss metric functions in the deep model to suppress the training errors and verify our training scheme by quantitative analysis. To address the potential boundary artifacts, we propose three kinds of refinement schemes and make a comparison on their performances. In the experiment section, we not only validate the reasonableness of our network parameters setting, but also conduct further exploration and analysis. Moreover, our experiments demonstrate this model can output more visual satisfactory colorization and obtain a better quantitative result when compared with the state-of-the-art methods. Last, we prove our method has extensive application domains and can be applied to stylistic colorization.