Research And Implementation Of Image Semantic Segmentation Based On Embedded Platform

Image semantic segmentation is a dense prediction task that aims to distinguish pixels belonging to different semantic categories in an image.As the cornerstone technology for intelligent machines to understand the real world,semantic segmentation has a good development prospect in the fields of medical image diagnosis,automatic driving,and security monitoring.At present,most of the semantic segmentation algorithms based on deep learning aim to obtain high precision.The huge network structure needs to consume a lot of computing resources,which results in slow inference speed,making it difficult for real-time applications.This thesis aims to study lightweight semantic segmentation algorithms suitable for embedded platforms to achieve the balance of accuracy and speed of semantic segmentation on intelligent edge devices.Semantic segmentation is essentially a process of encoding and then decoding the input image.This thesis is based on this process to carry out the research of semantic segmentation algorithm.Taking into account the limited hardware resources of the embedded platform,the encoder in this thesis chooses the lightweight classification network Mobile Net V2 as the backbone network for extracting features,and uses continuous downsampling and shrinking the number of feature channels to further reduce the amount of network parameters and complexity.So as to restore the edge details of the object lost by the encoder during the downsampling process,the decoder uses the skip connection method to fuse the feature output of different stages in the network by splicing,making full use of the deep high-level semantic features of the network and the high-resolution spatial detail of the shallow network.Aiming at the problem of object multi-scale and rough edge segmentation,this thesis proposes a hybrid attention module(Hybrid Attention Module,HAM)based on self-attention mechanism.This module fuses features of the two dimensions of channel and space to model dense context information,thereby strengthening the semantic expression of the output feature map and improving the segmentation accuracy of the model.This thesis trains and tests the proposed semantic segmentation model on the Cityscapes dataset.In terms of accuracy,the average m Io U ratio is 74.4%;in terms of speed,t it takes 23 ms for a single GTX 1080 Ti graphics card for forward inference,and 100 ms for Jetson TX2 embedded platform.In order to verify the effectiveness and practicability of the model,this thesis deploys the proposed semantic segmentation network on an embedded platform and applies it to different practical scenarios.For the application scenarios of roadside parking lot intelligent management and charging,this thesis makes improvements on the basis of existing work,integrates semantic segmentation functional components in the parking system,and makes up for the lack of target detection algorithms in the original system through fast semantic segmentation,improving the overall accuracy of the system.For the application scenario of smart city intersection supervision,this paper designs and implements a visual analysis system based on Py Qt5.With the semantic segmentation model proposed in this paper as the core,it intelligently perceives the on-site environment and extracts real-time information to assist managers in the analysis and monitoring.