基于光学神经网络的语义分割

Semantic Segmentation Based on Optical Neural Networkn

摘要: 语义分割是计算机视觉领域的关键挑战之一。目前，语义分割主要是基于卷积神经网络(Convolutional neural network, CNN)的各种深度学习算法来实现。然而，受限于传统的冯诺依曼架构，随着深度学习算法复杂度的不断增加，基于电子计算机的神经网络面临着算力，数据吞吐率和能耗的诸多限制，由于加工工艺已经接近了后摩尔时代的极限，进一步提升性能变得非常困难。本文提出了一种基于光学神经网络(Optical neural network, ONN)的图像语义分割方法。光学技术具有信息计算传输速度快、信息携带能力强、抗干扰性好等特点，在光神经网络中，矩阵乘法以光速进行并行计算，并且仅需极低的能量消耗，有效解决人工神经网络中面临的复杂的矩阵乘法问题。此外，基于存算一体架构，光神经网络的调制层既可以存储学习的参数，又参与了光学计算，从而避免了计算机面临的数据读取瓶颈。我们所设计的光学神经网络在Portrait 2000数据集上进行训练和测试，并在Person Correlation Coefficient (PCC)指标中达到0.8，有着良好的语义分割的效果。

Abstract: Semantic segmentation is one of the key challenges in the field of computer vision. At present, various deep learning algorithms based on convolutional neural network (CNN) are employed in this task. However, owing to the serial nature of von Neumann architectures, the electronic hardware platforms now confront significant challenges of computing speed, data throughput, and energy consumption, and the performance growth has become unsustainable due to the manufacturing process of electronic transistors approaches its physical limit. In this paper, an image semantic segmentation method based on optical neural network (ONN) is proposed. ONN has the characteristics of fast computing speed, low computing latency, and strong information carrying capacity. For the optical neural networks, matrix multiplication can be parallelly calculated at the speed of light with more energy-efficient, which effectively solves the complex matrix multiplication problem in artificial neural networks. Furthermore, as a computing in memory architecture, the modulation layers of ONN not only store the learned parameters but also participate in the optical calculation. The network is trained and tested on the Portrait 2000 dataset and achieved 0.8 in the Person Correlation Coefficient (PCC) index, showing good segmentation effect.