Single-Pixel Moving Object Classification with Differential Measuring in Transform Domain and Deep Learning

Abstract

Due to limited data transmission bandwidth and data storage space, it is challenging to perform fast-moving objects classification based on high-speed photography for a long duration. Here we propose a single-pixel classification method with deep learning for fast-moving objects. The scene image is modulated by orthogonal transform basis patterns, and the modulated light signal is detected by a single-pixel detector. Thanks to the property that the natural images are sparse in the orthogonal transform domain, we used a small number of basis patterns of discrete-sine-transform to obtain feature information for classification. The proposed neural network is designed to use single-pixel measurements as network input and trained by simulation single-pixel measurements based on the physics of the measuring scheme. Differential measuring can reduce the difference between simulation data and experiment data interfered by slowly varying noise. In order to improve the reliability of the classification results for fast-moving objects, we employed a measurement data rolling utilization approach for repeated classification. Long-duration classification of fast-moving handwritten digits that pass through the field of view successively is experimentally demonstrated, showing that the proposed method is superior to human vision in fast-moving digit classification. Our method enables a new way for fast-moving object classification and is expected to be widely implemented.