3D Point Cloud Semantic Segmentation System
Kuan-Yu Liao, You-Sheng Xiao, Yu‐Cheng Fan · 2022
With the continuous progress of science and technology, artificial intelligence and big data are also constantly evolving. In artificial intelligence, the 2D image recognition of deep learning is relatively mature, but the 3D point cloud image is still in development. Because the data of 3D point cloud is larger than that of 2D image, and the scanning method is different, which affects the point cloud image to generate different data types. For various reasons, the development of 3D point clouds still requires us to make more efforts to study how to improve them. With the continuous expansion of the data size, the computing performance becomes more important, and in deep learning, the convolutional layer occupies a very important position. If the performance can be improved in the convolutional layer, it will definitely bring good benefits to artificial intelligence, so this paper will propose the use of efficient reuse, depthwise separable convolution, quantization, and Winograd algorithm to accelerate the operation and reduce the weight. Finally, according to the digital integrated circuit design process based on the standard cell library, the convolution layer operation in the neural network is realized.