Research On Real-time Semantic Segmentation Of Mobile Image Based On Deep Learning

In recent years,deep learning has developed rapidly.Image semantic segmentation is a hot topic in the field of deep learning.Image semantic segmentation is of great significance to the development of computer vision and practical applications,especially in the field of autonomous driving.Image semantic segmentation plays a significant role in understanding the driving environment.With the popularization and application of various mobile devices such as mobile phones,the integration of deep learning on mobile devices and the effective application of convolutional neural networks have profound value.However,the mobile terminal’s memory resources and hardware conditions are still facing great challenges for the rapid application of convolutional neural networks.To better apply the convolutional neural network visual model,it must balance its accuracy and the performance constraints of mobile runtime performance.In order to solve this problem faced by the visual perception module in the mobile terminal field,this paper analyzes and researches the existing algorithms aiming at real-time online segmentation of mobile devices.Finally,based on the classic Deep Lab V3 model,and under the premise of ensuring segmentation accuracy,a new lightweight network structure is proposed mainly to optimize the running speed of the convolutional neural network model.The network structure designed in this paper maintains a similar main structure to Deep Lab V3,that is,the network model is mainly composed of the coding part structure.The coding is the feature extraction part.The Mobile Net V2 network structure is improved first.The original non-linear activation function is replaced with a new Swish activation function for accuracy compensation.The improved lightweight model Mobile Net V2 structure replaces the original Deep Lab V3.The feature extractor finally makes the improved Deep Lab V3 network model maintain a certain degree of accuracy,the parameter amount and the computational complexity are greatly reduced,and the calculation speed is significantly improved.In order to test the generalization ability of the model,the model was trained using two different Pascal VOC 2012 and Cityscapes datasets.In this paper,the Deep Lab V3 + algorithm with the best effect of Deep Lab series and the improved Deep Lab V3 algorithm are compared and tested.It is found that the improved Deep Lab V3 network can significantly improve the running time of the model without losing much accuracy.By obtaining an improved deep learning model and applying it to mobile robots for experimental verification,real-time semantic segmentation of indoor scenes can be completed,the basic attributes of indoor scene objects can be obtained,and advanced conditions for mobile robot map construction and autonomous navigation can be created.The experimental results show that the improved lightweight model of Deep Lab is the best for real-time image semantic segmentation for mobile end,from the comprehensive performance of segmentation effect and running speed.