Research On Image Semantic Segmentation Algorithm Based On Deep Learning And Its Application

Image semantic segmentation is an important method for computer to understand and handle images.It is widely used in medical image processing,satellite geographic information system,robot vision and so on.Therefore,this technology has attracted more and more attention from scholars.Deep learning is a discipline that further imitates the human nervous system.The key content is to use the convolution operation to integrate the input data,extract more representative feature data,find the data distribution from this high-dimensional feature,and then complete a variety of tasks.Benefiting from the excellent computing power of convolutional neural network,method based on deep learning can process large quantities of data while ensuring better segmentation results,which brings a new method for image semantic segmentation.However,the current deep learning solutions still have many shortcomings.The neural network needs to be trained for a long time before use,and the effect of training the network depends on the network structure and parameter settings;in addition,the training of the neural network needs to use a large amount of finely labeled data as labels,and this fine labeling Data is not easy to obtain,not only requires a large number of professionals and tools,but also consumes a lot of time and energy.Based on the above reasons,this article mainly did the following work:(1)For the feature map does not contain long-distance pixel dependence.This paper proposes a multi-threshold probability segmentation method based on feature fusion,which can improve the accuracy of image segmentation.Based on Mask Scoring R-CNN,this algorithm proposes a feature fusion method that can integrate high-dimensional and low-dimensional features,and adds it to the feature pyramid network to optimize feature extraction;in addition,we add to the network Multi-threshold segmentation architecture to improve the screening of samples,and at the same time use probability models to further optimize the segmentation branch of the algorithm to improve the accuracy of the algorithm's segmentation of image edges.Experiments show that this algorithm improves the final segmentation result.(2)For weakly supervised networks,more spatial location information is needed and weakly supervised labels contain less information.This paper proposes a weakly supervised image semantic segmentation algorithm based on attention mechanism,which can use bounding boxes as labels to achieve semantic segmentation of images.In this method,an attention model that can aggregate the dependencies between the spatial domain and the channel domain is proposed to enhance feature extraction.The basic steps of the algorithm are divided into two steps,generating temporary semantic segmentation labels and training neural network models,and then iterating these two steps.Our experiments on PASCAL VOC 2012 prove that this algorithm is effective and its performance is improved,which is close to the performance of the corresponding fully-supervised network.(3)Using the Python language and the Pytorch deep learning framework,the two algorithms proposed in this paper have been verified for practical applications in urban landscape segmentation and medical image segmentation scenes.The results show that our algorithm can reach a higher level.