Research On Light Weight Semantic Segmentation Based On Convolutional Neural Network

Semantic segmentation is an important research field of computer vision.It has important applications in the automatic driving of vehicles,drones,intelligent robots,etc.It has become one of the most popular fields of computer vision in recent years.Although many semantic segmentation methods have been proposed at home and abroad,due to factors such as large input image resolution and computer performance limitations,traditional semantic segmentation has defects such as excessive calculation and inability to achieve real-time segmentation speed.On the other hand,real-time semantic segmentation often cannot accurately segment small objects such as people,bicycles,and traffic lights in the input image due to the small network structure and severe loss of resolution.Based on the convolutional neural network model in deep learning,this paper conducts in-depth analysis experiments to solve the problems of the slow speed of traditional semantic segmentation and the difficulty of real-time semantic segmentation for small target object recognition.And actively explored and proposed improved methods,which significantly improved the recognition accuracy of the network for small target objects to meet the requirements of real-time segmentation speed.The main research work of this paper is as follows.(1)A two-way multi-scale network is proposed to aim at the slow reasoning speed of the traditional semantic segmentation network due to its huge network and too many parameters.The network includes a high-resolution network and a low-resolution network.The high-resolution network is composed of a lightweight network,which receives the input image of the original image resolution and outputs high-resolution semantic information;the low-resolution network is composed of a normal network.But it only accepts 1/2 resolution input image,and the resolution drops rapidly,so it can extract deep context information with low computation.In addition,the auxiliary segmentation Loss is added to assist network training during training,and the network accuracy is further improved without increasing the amount of calculation during inference.In the end,71.1% m IoU was obtained on City Scapes,a common dataset for semantic segmentation.(2)Aiming at the problem of difficult segmentation of small target objects caused by simple network structure and weak expressive ability in real-time semantic segmentation.We propose an lightweight convolutional neural network based on fusible convolutions.By introducing the fusing operation of convolution,multiple convolution operations are used to extract features of different scales during training.They can be fused into a single convolution during inference to ensure the inference speed.The ultra-lightweight network redesigned based on this structure further improves the high-resolution features of the two-way multi-scale network to 1/4,which makes the network significantly improve the segmentation accuracy of small target objects.In addition,the feature fusion module that can be redesigned by fusible convolution and dilated convolution further enhances the network accuracy at almost the same inference speed.On the dataset City Scapes,we achieved 74.5% m IoU.