Research On Scene Segmentation Based On Convolutional Neural Network

In recent years,deep learning algorithms based on convolutional neural networks have achieved great success in a variety of computer vision tasks,playing an important role in the field of image processing.Scene segmentation is an important research direction in the field of computer vision.It aims to segment different objects and regions in the image according to their attributes under the supervision of pixel-level labels,and assign category labels to them.At present,image segmentation algorithms based on deep learning have important roles and are widely used in fields such as autonomous driving,modern image medicine,and industrial inspection.As an important research branch in image segmentation,the scene image segmentation task is a challenging visual task due to the many and complex targets in the scene image and the changing scene styles.On this basis,this paper focuses on the scene image segmentation technology based on convolutional neural network.Since the image segmentation algorithm based on convolutional neural network can automatically obtain the semantic features of different layers from shallow,middle to high layers in the image,the accuracy of image segmentation has been qualitatively improved.However,the traditional segmentation algorithms still have the problems of loss of edge details in complex backgrounds,and low accuracy of boundary segmentation.This paper firstly revises on the basis of the classical fully convolutional neural network,and proposes a fully convolutional neural network model that merges context information.The model is presented in an encoding-decoding structure,in which the encoding module is composed of a backbone module and a feature fusion module.The backbone module is the full convolutional neural network.It is used to extract the semantic features of different levels of the image,and the feature fusion module is used to fuse the shallow features with the deep features so that there is more richer context information when encoding,and it can also avoid the gradient disappearing during back propagation.The feature fusion module uses two parts: dilated convolutional layer and residual gate structure,which uses dilated convolutional layer to retain shallow detail features.Then use the residual gate structure to filter the shallow details features,enhance the more significant features,and weaken the weak response features.The decoding module applies the deep features of the convolution part to the deconvolution network part through continuous upsampling.Experimental results show that the model has improved segmentation accuracy on the PASCAL VOC 2012 dataset compared with other methods,and has better detail retention and grasp of the overall shape of the target.Since scene image segmentation is widely used in fields such as automatic driving and robot sensing,in network construction and practical applications,network complexity and algorithm efficiency are important factors that cannot be ignored.However,at this stage,the network will adopt a deeper network layer and a larger receptive field in pursuit of higher segmentation accuracy,which will slow down the calculation speed of the neural network and is not easy to segment in real time.This paper proposes a lightweight neural network model based on multi-scale feature fusion.The model uses a lightweight network backbone and data-dependent upsampling module to compress network parameters and increase network speed.And It uses feature aggregation and position attention modules between multi-layer backbone networks to effectively retain spatial information and context information to ensure the segmentation accuracy of the model.Experimental results prove that the model has good network estimation speed on the Cityscapes dataset while ensuring segmentation accuracy.