Abstract:

The physical architecture of optical convolution restricts its capacity to capture multi-scale features from targets, thus impeding the precision of network recognition. In this work, we propose a multi-scale optical convolutional neural network (MS-OCNN), which uses convolution kernels with different resolutions in the convolution layer to extract different scale features, along with attention mechanism and residual structure to analyze features. By separating the training and inference platforms of the network, we facilitate the electronic training of model parameters on a computer and the optical deployment on a system equipped with a spatial light modulator, enabling efficient target classification. The proposed MS-OCNN exhibits 1% to 3% improvement in classification performance on the modified national institute of standards and technology (MNIST) and Fashion-MNIST datasets compared to single-scale optical inference models. Online experimental systems in real-world scenarios have validated the target recognition capabilities of this method, which yielded classification accuracies of 97% and 87% on the MNIST and Fashion-MNIST datasets, respectively. This work enhances the feature acquisition capabilities of optical convolutional networks, elevates network recognition accuracy, and significantly propels the application of optical computing in domains such as guidance systems, autonomous driving, and robotics.

Key words: optical computing, multi-scale features, target classification, guidance system