Image-free object detection from single-pixel measurements with learned binary patterns

Single-pixel imaging (SPI) is a novel technique that captures 2D images using a programmable spatial light modulator (SLM) and a single-pixel detector, instead of conventional 2D array sensors. The image can be reconstructed from the modulation patterns and corresponding 1D bucket measurements. Conventional object detection is performed after a reconstructed image with high fidelity is available. In this paper, an image-free object detection method using the single-pixel measurements is proposed. We designed and trained an end-to-end convolutional neural network to encode and decode a scene for image-free object detection application. The performance of the proposed method is demonstrated using part data from the Voc dataset, which achieves a detection accuracy of 35.41% mAP at a sampling rate below 1.6%.