Image-free object classification from compressive single-pixel measurements

In conventional approaches, object classification requires image acquisition before any content analysis can be performed. In contrast, single-pixel imaging (SPI) captures image information indirectly, encoding it in a transformed and compressed representation that retains essential features of the scene. The original image can be reconstructed from SPI measurements, however doing so requires solving an ill-posed inverse problem, which is computationally expensive and often impractical. To address this challenge, we propose a deep learning-based approach that directly classifies image content from SPI signals, eliminating the need for image reconstruction. We investigate two configurations: one that uses a fixed, deterministic sampling matrix based on low-frequency elements of the discrete cosine transform (DCT); and another that jointly optimizes both the sampling matrix and the classification model, enhancing the efficiency of feature encoding and extraction. By operating directly on measured SPI data, it significantly reduces computational costs and memory requirements compared to conventional classification methods that rely on image reconstruction. This makes it particularly well-suited for resource-constrained applications such as remote sensing systems or embedded vision.