A Semantic Segmentation Algorithm Using Multi-scale Feature Fusion With Combination Of Superpixel Segmentation

With the rapid development of digitization and artificial intelligence,data has become an important resource in the computer age,and the importance of image and video data has become increasingly prominent,so the field of computer vision research has attracted more and more attention from researchers.Image semantic segmentation is one of the most important basic research issues in the computer vision.Its research results are widely used in real life and have important academic and practical values.With the development and popularization of deep learning,image semantic segmentation methods based on deep learning have achieved substantial breakthroughs,but further research is needed on issues such as image edge detail segmentation and high-and low-level image feature fusion.Traditional image semantic segmentation technology usually uses the similarity between pixels to divide each pixel in the image,which cannot meet actual needs in terms of segmentation accuracy and efficiency.Aiming at the problems of image edge detail segmentation and high and low-level image feature fusion,we propose two methods: a multi-scale feature fusion image semantic segmentation algorithm combined with superpixel segmentation and an improved image semantic segmentation algorithm based on kernel shared convolution.For the multi-scale feature fusion image semantic segmentation algorithm combined with super-pixel segmentation,firstly,the deep learning model is used for image feature extraction,and then the joint cross stage partial multiscale feature fusion module implemented in this paper is used for multiscale feature extraction and fusion,then we use the decoder to obtain rough semantic segmentation results.In addition,a jump connection structure is added to the network to enhance the learning ability of the network.Finally,the superpixel edge optimization algorithm module is used to rough the network prediction results.For the improved image semantic segmentation algorithm based on kernel shared convolution,firstly,a fully connected deep dense network is proposed as the basic network of the algorithm based on the Dense Net.Then we proposed the cross stage partial shared atrous spatial pyramid pooling module to enhances the image feature fusion and reduces the computational complexity.In the experiment process,we adopt the commonly used datasets of image semantic segmentation such as PASCAL VOC 2012 and Cam Vid.The effectiveness and robustness of the algorithms proposed in this paper are proved by analyzing and comparing the experimental results on the two datasets.In the experiment,we also compare with some excellent image semantic segmentation algorithms,the experimental results show that the improved algorithm proposed in this article not only performs well in subjective observation,but also has a certain improvement on the recognized evaluation standards PA and mIoU.