Research And Implementation Of Semantic Segmentation Of Indoor Scene Image Based On Embedded Platform

Image Semantic Segmentation is an important and difficult visual task in the field of computer vision.The goal is to complete the classification of each pixel.Its unique task nature makes it widely used as a pre-task in many scenarios,such as autonomous driving,indoor service robot scene understanding,intelligent matting,face segmentation,and augmented reality.At present,image semantic segmentation mainly uses convolutional neural network for semantic segmentation.However,the existing semantic segmentation network based on deep convolutional neural network has deep layers,complex structure,and slow inference speed.If high-precision real-time inference is required,a large amount of computing power is required,and the hardware requirements of the computing platform are high.Real-time application scenarios such as autonomous driving,service robots,and augmented reality computing devices are often not as good as high-performance and high-power GPU servers in terms of hardware computing power and power consumption.In response to this problem,this paper,based on the deep convolutional neural network,starts with the semantic segmentation network and model architecture,and conducts in-depth research on efficient real-time indoor scene semantic segmentation methods.The specific research work is as follows:First,starting from the design of a lightweight model,this paper proposes a lightweight codec segmentation network for real-time semantic segmentation,called the Lightweight Cascaded Segmentation Net(LCSNet).The method in this paper uses a lightweight encoder-decoder network to solve real-time semantic segmentation tasks.The encoder proposes a novel depth separable hole convolution residual coding module,introducing a path aggregation module and a channel attention mechanism,and the decoder uses lightweight upsampling plus multi-level feature fusion to extract efficient semantics feature.The experimental results show that,compared with the current advanced deep convolutional network model,the parameter of the method in this paper is only 3.1M,and the inference speed of 86 FPS is reached on Titan XP GPU.On the supplemented scene semantic segmentation datasets NYUDv2 and Cam Vid,segmentation accuracy of 73.7% and 59.9% were achieved respectively.Second,for embedded platforms of different architectures,this article uses convolutional neural network acceleration calculation methods to accelerate the optimization of the embedded platform neural network segmentation algorithm.Experiments show that after general matrix calculation optimization and Winograd algorithm speed up the convolution calculation,the algorithm's inference speed on the embedded platform is increased by more than 200%.At the same time,model compression technology is used to optimize the volume of the model.After optimization,the model volume is only 1.2M,which greatly reduces the memory usage of the embedded platform.Third,facing the deployment of semantic segmentation algorithms for indoor scenes,this article will deploy and test the optimized LCSNet algorithm on the embedded GPU Xavier NX platform and the mobile phone equipped with Qualcomm 865 embedded CPU.The method proposed in this paper has been deployed and optimized by Tensor RT to achieve a speed of 52 FPS on the Xavier NX platform,and a speed of 7FPS on the mobile phone.The experimental results show that the method proposed in this paper can complete real-time semantic segmentation of indoor scenes on the embedded GPU side,and can complete the matting application on the mobile phone side,and face segmentation and other applications that have less stringent real-time requirements.