Aerial Image Classification through Thin Lensless Camera

FlatCam - a lensless camera, is characterized by its thin form factor, flexibility, and low power consumption. These are the ideal characteristics for imaging system in an Unmanned Aerial Vehicle. FlatCam captures the scene as sensor measurements and requires computational algorithms to reconstruct the scene. In this paper, we propose a reconstruction-free frequency domain deep learning based aerial image classification system. The system directly learns from sensor measurements thereby eliminating the computational cost associated with image reconstruction. The proposed approach computes the Discrete Cosine Transform of the sensor measurement and organizes it into subbands to learn an effective embedding for classification with a two-stage training using softmax and triplet loss. A triplet sampling strategy that uses label splitting is used for triplet loss learning. Our system is lightweight and the simulation results demonstrate its effectiveness.