Semantic Segmentation Based on Multi-Channel All-Optical Convolutional Neural Networks

目前，光学神经网络作为一种全新的人工智能技术，以其高算力和低能耗得到了广泛的研究和应用。但是目前常规的光学神经网络结构简单，缺乏对复杂任务的处理能力。文章提出了一种全新的多通道光学卷积神经网络(MC-OCNN)架构，来处理机器视觉中的语义分割问题。MC-OCNN由透镜相位、随机相位，以及光栅相位构建。透镜相位用来实现光学卷积操作，随机相位用来组成光学卷积核，以及光栅相位用来实现多通道架构。MC-OCNN具备多层次结构，由光学卷积层以及光学全连接层级联而成，以光速并行进行卷积计算，并且实现了多维度提取物体不同特征，以及多层次提取物体特征的功能。我们将MC-OCNN应用在人像分割任务以及飞机检测任务中，取得了良好的结果，为其进一步在机器视觉、自动驾驶、智慧医疗等领域的应用奠定了基础。Currently, optical neural networks (ONNs) have attracted wide research and application due to their high computational power and low energy consumption. However, the structure of conventional ONNs is simple, and their ability to handle complex tasks is very limited. Here, we propose a novel multi-channel optical convolutional neural network (MC-OCNN) architecture to address the semantic segmentation problem in machine vision. The MC-OCNN is composed of the lens phase, random phase, and grating phase. The lens phase is used to implement the optical convolution, the random phase is used to form the optical convolutional kernel, and the grating is used to construct the multi-channel architecture. MC-OCNN has a multi-layer structure formed by cascading optical convolutional layers and fully connected optical layers, and it also could parallelly perform convolution computations at the speed of light. Besides, it achieves the extraction of different features of the objects, as well as the multi-level extraction of object features. The MC-OCNN is applied to portrait parsing tasks and aircraft segmentation tasks, achieving promising results that demonstrate great potential for application in fields of machine vision, autonomous driving, smart healthcare, and so on.