Research And Implementation Of Convolutional Neural Network Algorithm Based On Level Set In Semantic Segmentation

With the rise of artificial intelligence,deep learning is one of the hottest research directions today,and it facilitates people in many applications.However,there are many problems in the classic convolutional neural network.In the process of feature extraction,the resolution will continue to decrease,resulting in low segmentation accuracy of the last layer features on the edge of the object,which limits the detection accuracy of the target.In view of the above problems,this paper proposes a convolutional neural network algorithm combined with the Level Set method to train a complete image segmentation network,and finally deploy it on a web server to produce a small semantic segmentation system.The main research content is divided into the following parts:CNN-based segmentation networks produce low-resolution outputs with rich semantic information,so spatial details(e.g.,small objects and fine boundary information)of segmentation results are inevitably lost.To address this problem,driven by a variational approach to image segmentation(i.e.,level set theory),a novel loss function called the level set loss function is proposed by combining the traditional level set method with a deep learning architecture,It aims to refine the spatial details of the segmentation results and improve the segmentation accuracy.To handle multiple classes in an image,the ground-truth image is first decomposed into multiple binary images,each binary image consists of background and regions belonging to one class.Then the level set function is transformed into a class probability map and the energy of each class is calculated,and the network is trained to minimize the weighted sum of level set loss and cross entropy loss.The proposed level set loss improves the spatial details of the segmentation results.Experiments and verifications on the dataset illustrate the feasibility of the application of the level set method.U-Net uses the skip connection scheme to model global multi-scale context information.Although the structure is simple,there are deficiencies: the feature sets of the encoder and decoder stages are incompatible,which will have a negative impact on feature fusion.In some data sets It is worse than the effect without skip connection structure.Based on the above problems,the ULevel Net-Attention model is proposed,which introduces the attention mechanism from the perspective of the channel,replaces the original skip connection,solves the problem of semantic gap,realizes accurate image segmentation,and captures more boundary information in the image.Experiments show that the improved network exhibits more accurate and stable segmentation performance.In this paper,the performance of the segmentation method is verified and applied through the PASCAL VOC dataset and ISBI Challenge 2012.The experimental results show that the accuracy of the algorithm in this paper can be further improved in object segmentation.The client-side semantic segmentation system was built by using the Spring Boot and Flask framework,which realized the semantic segmentation of the input image on the browser,and added Redis to speed up the system response and improve the practicability of the deep learning network model.