Research And Implementation Of The Costume Generation For Scratch.3.0

Scratch is a creative and visual tool to teach programming for young students.It had already been widely used in recent years to help teenagers learn how to program.Before programming with Scratch,users usually need to select or draw a sprite that can use for the project later.However,there are no options in Scratch that can translate a sketch to an image though it has a paint-board that allows users to design a costume for sprites.Because of this reason,we intend to include an image style translation mechanism with Scratch 3.0.However,including an image style translation mechanism with Scratch 3.0 is a challenging process.Due to solve the complexity,we divide our task into three parts:1)translating sketch-to-image 2)transfer neural-style-to-image 3)integration the two functions with Scratch 3.0.First,we design a novel Conditional Generative Adversarial Networks,which not only can translate an image from a sketch but also remove the unwanted artefact.Unlike including a U-Net generator architecture,we use a modified MultiResU-Net generator that uses the residual path as a skip connection.By using this modified architecture,the number of parameter of the generator and period of training time decreases dramatically.Second,in order to make the image achieve cartoon style effect,we develop an end-to-end deep learning method.We propose to combine the CartoonGAN with the Dense-layer to generate cartoon-style images.In particular,a new end-to-end photograph cartoonization method called Dense-CartoonGAN is proposed,which can map between real-life representations and the cartoon image.End-to-end learning is achieved by replacing the residual block with the dense block in the generator network.The advantages of the Dense-CartoonGAN is strengthening propagation,encouraging feature reuse,and substantially reduce the number of parameters.Therefore,Dense-CartoonGAN obtains significant improvement over the CartoonGAN while requiring less computation to achieve high performance.Finally,we develop an integration application for Scratch 3.0 with image style translation.Since our proposed image style translation generates better quality outputs,using an application like this gives the user a satisfactory completion.