Shift-invariant image classification using a bicolor shadow-casting incoherent optical system

In this study, a shift-invariant optical pattern classification system is proposed. Optical machine learning systems have been widely studied as processors with massive parallel computing and low power consumption. Conventional optical systems used for pattern classification require diffractive optical elements with microscale surface structures or lens systems. The target images and optical elements require precise alignment. The proposed system comprises a liquid-crystal display, a target image, and an image sensor. Despite not requiring complex optical elements or alignment precision, distorted patterns are classified based on linear discriminant analysis (LDA), and high classification accuracy is maintained irrespective of the position of the target image. Classification accuracy and shift invariance were validated experimentally using a handwritten digit image dataset.