Research On Automatic Image Coloring And Animation Face Generation Based On Lightweight Neural Network

Color processing of images has been widely used in people's life and scientific research.Following the advancement of artificial intelligence,there are lots of results on the prediction of grayscale images using deep learning techniques.As more and more services require coloring effects with higher quality,some schemes have explored models.These models were more complex and deeper levels of network structure to cope with the corresponding demands.As a result,it has a very large model parameter space which cannot be adapted to environments with limited computational power or memory.Furthermore,some deep learning-based models fail to achieve color images with complex semantics.The main reason is that they don't efficiently locate and learn meaningful instance-level semantics.So,it is of great interests to utilize lightweight neural networks with instance-level semantic information for grayscale image coloring.This thesis researches the grayscale image coloring problem based on lightweight convolutional neural networks with weak semantic segmentation.The main results are as follows.One lightweight coloring model is designed with global features.The other coloring is designed model with weak semantic segmentation.By constructing a new loss function,a lightweight coloring model with global features is established by using the depth-separated convolution and the global features module.the parameter of model is 8.38 M and the computational complexity is 1.36 G.The proposed lightweight automatic coloring algorithm has better performances compared with previous algorithms.The module of weak semantic segmentation is introduced based on the aforementioned coloring model with global features.Interesting,the instance-level semantic information can be extracted from improved model.Moreover,a PSNR of 23.85 d B can be obtained in the VOC2012 test set experiments.This thesis shows that the proposed coloring algorithm is reasonable.A GAN-based anime face generation model is designed,and the model is compressed by using compression technique.The model is constructed by combining the Ada LIN normalization function and attention mechanism in U-GAT-IT,and pruned by the Slimming technique.These guarantee the effects of generating anime faces with less computational complexity.Finally,the face 3D reconstruction problem will be addressed by presenting the test experimental results,the proposed schemes reconstruct 3D real faces from sketched faces and 3D anime faces from real faces.