Object Classification using Deep Neural Networks from Coded Diffraction Patterns

Deep learning (DL) has been broadly used in computational imaging tasks, such as object classification. However, traditional DL approaches for object classification are based on intensity images sensed with linear propagation systems. Since optical sensors can only measure the intensity underlying in the complex optical field, these linear systems capture only magnitude information, while losing the phase information. In order to recover the phase information, diffractive optical imaging systems that modulate the phase information and capture them as coded diffraction patterns (CDP) have been proposed. These CDP are acquired by including coding masks that modulate the optical field before being recorded in the sensor. This work proposes an object classification approach by using DL from CDP. The proposed scheme consists of three main stages: an acquisition layer that simulates the CDP capture process; an initialization step that applies by using separately two methods such as a back propagation operator of the optical field and a filtered spectral algorithm; a deep neural network that performs the classification task. The simulation results over the MNIST and Fashion-MNIST datasets show that the inclusion of the initialization strategy in the general network setup improves the object classification from CDP performance compared to a classification scheme based on none initialization.