Trans-ONN Image Classification: Unveiling Translation Invariance in All-Optical Networks

Translation-invariant optical neural network for image classification

In practical applications, the misalignment of components and translation of inputs images greatly affects the classification performance of all optical Convolutional Neuronal Networks (CNNs). This paper proposes a free space all-optical CNN named Trans-ONN that accurately classifies images translated in horizontal, vertical or diagonal directions. Trans-ONN uses an optical motion pooling (OMP) layer to provide translation invariance by implementing different masks in Fourier plane. Trans-ONN uses global average pooling to improve translation invariance.