Understanding Cartoon Characters Through The Body Parsing Task Using Deep Learning

Cartoon characters are a form of art that is extremely difficult for machines to understand.The body parsing is a method similar to the finegrained semantic segmentation that can help machines to understand cartoon characters.While studies on the cartoon parsing task are very limited,research on naturalistic images are abundant.In this field,the leading research uses deep convolutional neural networks(DCNNs),however it demands a considerable supply of data to have great results.Because of the infinite amount of unique drawing styles and the sparsity of datasets,the complexity of cartoon character parsing is higher than the famous human parsing task.In this thesis,cartoon dogs are the instance type we decided to work on.We create a new dataset adapted to the cartoon dog parsing task composed of965 cartoon dog drawings.Each drawing has an meticulous annotation of its7 semantic body parts.We design a novel deep convolutional neural network(DCNN),named Dense Feature Pyramid Network(DFPnet),for cartoon dog body semantic segmentation.In order to create the best model for our task,we conduct extensive experiments on the latest semantic segmentation approaches.DFPnet surpasses state-of-the-art techniques of comparable work on our cartoon dog dataset.Our DFPnet attains 68.39% with m Io U,93.5%with Pixel Accuracy and 79.4% with Mean Accuracy on the test set.Our study could be a helpful foundation for following studies toward digital artwork understanding.