Research And Implementation Of Automatic Icon Creation System Based On Neural Network

When individual developers or small teams develop software,they often use the free resources of the Internet for the sake of cost,and the copyright of these resources is mostly unclear.These free resources play a role in promoting the early development of low-cost software,but when these software form a certain scale,they may face copyright disputes and pay huge inspection costs.In order to better solve these problems,a reasonable idea is to use the icon generation method based on neural network.Because a neural network can generate many pictures,it can save the cost of investigating and verifying the copyright of a large number of pictures.At present,there are many different technical routes about this,such as using the style migration algorithm to process the user's input images to get new images,or using the confrontation generation network to get the generated images through random seeds.The stability of these methods will decrease rapidly when they generate higher pixel images.In view of the multimodality and unique morphological characteristics of icon resources,this paper proposes a differentiable vector fitting algorithm,which makes the neural network better deal with vector graphics by representing the graphics as the combination of primitives.In order to improve the stability of the neural network,the algorithm also uses the confrontation generation network based on self encoder clustering.The experimental results show that,compared with the method of generating bitmap based on Countermeasure generation network,this scheme reduces the computational complexity and avoids the problems of unstable training and difficult convergence when generating high-resolution images.Based on the proposed vector fitting icon generation algorithm,this paper designs and implements a vector icon aided creation system.The system can produce icon resources according to the user's input category information and its own randomly generated seeds,and continuously obtain the icon feedback produced by neural network by walking in the latent space,so as to provide users with a variety of self-defined icons.