Research On Image Semantic Segmentation Algorithm Based On Convolutional Neural Network

Traditional image semantic segmentation techniques are limited by precision and speed and cannot meet complex scene requirements.With the development of deep learning technology,more and more researchers have begun to use the convolutional neural network to solve the semantic segmentation problem and promote the rapid development of semantic segmentation.Performance in terms of accuracy and speed goes beyond traditional segmentation methods.The image semantic segmentation algorithm classifies the input image at the pixel level.The goal is to achieve intensive prediction of each pixel of the image such that each pixel is labeled with the category of the corresponding object or region.From the segmentation results,the comprehensive semantic information of the scene can be obtained intuitively.In this paper,the semantic segmentation of images is deeply studied based on convolutional neural networks.In order to solve the complex scene problem in the interior,the multi-level and multi-scale feature fusion segmentation model based on the full convolutional neural network is firstly realized.Based on this,a multi-task that can achieve target classification,detection and segmentation simultaneously is proposed.The main work is as follows:The core components and principles of the structure of convolution neural network and full convolution neural network are studied in depth,including basic feature extraction network,down-sampling,up-sampling,transposition convolution and other operations.Then,the segmentation framework is determined based on this,and image semantics segmentation is realized.Aiming at the problem that the original segmentation algorithm is not accurate in the indoor complex road environment,an improved full convolutional neural network semantic segmentation algorithm is designed.The whole framework is divided into two stages: encoder and decoder.The task of the encoder is to extract the feature information of the input image and get the feature map.The decoder is to reconstruct the lost spatial information by sampling on the feature map,and ultimately achieve the segmentation of the image at the pixel level.The shallow features extracted from the network are merged with the deep abstract features,and multi-scale information is added,which makes the reference information of the network final prediction segmentation more comprehensive.By verifying on the experimental scene dataset,the convergence speed and accuracy of the algorithm are improved..A multi-task segmentation algorithm based on full convolutional neural network is proposed to solve the indoor complex environment segmentation.By adding the target detection branch to the segmentation algorithm,the task of combining object classification,detection and segmentation is innovatively realized,and the network is further optimized.By improving the network loss function,the algorithm is improved using deeper feature extraction networks and model optimization algorithms,and Transfer learning ideas are used to adjust the training.Verification is done on the MS COCO dataset and the indoor actual scene dataset.The experimental results show that the proposed algorithm has good robustness in indoor complex segmentation scenarios and improves the average accuracy..