Research And Implementation Of Sky Style Conversion Algorithm Based On Deep Semantic Segmentation

With the rapid development of science and technology,the sky style conversion application can meet the user’s pursuit of beauty,and you can enjoy a variety of sky beauty without leaving home.The development of sky style conversion application in today’s era of spiritual abundance is of great significance.However,there are some problems in the current technologies in this field.For example,the roughness of the sky area obtained by the traditional segmentation method leads to distortion in the sky style conversion application,and only use the style transfer method cannot achieve the problem of only converting the style of the sky area.To solve the above problems,this paper proposes an efficient and widely used sky style conversion algorithm.In order to facilitate the research and implementation of the sky style conversion algorithm,we constructed an accurately labeled sky segmentation dataset,which is specially used for training and testing the algorithm model proposed in this paper.On this basis,this paper proposes a sky style conversion algorithm based on deep semantic segmentation.Specifically,the semantic segmentation module not only maintains the high-resolution branch in the whole network,but also performs downsampling to extract the deep semantic information of the image,so as to achieve the fusion of feature information across resolutions;In terms of feature extraction,a multi-branch attention mechanism is introduced to enhance the feature capture capability of the network from both channels and convolution kernels;the refinement segmentation module uses the residual network structure and pyramid pooling to refine the edge of the sky segmentation mask;According to the size of the sky segmentation mask and the target sky color parameters,a sky style conversion and fusion module is designed to complete the sky style conversion task.The results of the comparison experiments with traditional semantic segmentation methods and the ablation experiments on the sky segmentation dataset show that the algorithm proposed in this paper achieves 83.68%mIoU and 46.31%mIoU on Cityscapes and ADE20K datasets respectively,and achieves the best performance of 97.89%mIoU on the sky segmentation dataset.Finally,based on the above algorithm,a prototype system of sky style conversion is designed and implemented to meet practical application requirements.