Task-driven single-pixel salient object detection via deep semantic compression

Single-pixel imaging (SPI) is a cost-effective computational imaging technique in special spectral bands. Traditional SPI-based sensing generally follows an “imaging-first” strategy, where reconstructed images may contain redundant information, such as irrelevant backgrounds. Consequently, this redundancy incurs additional hardware and computational costs during data acquisition, transmission, and storage. To mitigate this waste of system resources, performing advanced sensing tasks directly from 1D measurements is crucial. Therefore, we propose a task-driven, end-to-end single-pixel salient object detection (SOD) method that achieves accurate saliency detection directly from a small number of measurements, without explicit and complex image reconstruction. Specifically, the framework models the physical sampling process of SPI through a learnable task-aware sampling encoder, which modulates the target scene to obtain a compressed measurement vector. Subsequently, this vector is mapped into a 2D spatial representation via a fully connected layer and fed into the back-end network for pixel-level prediction. By jointly optimizing the learnable sampling layer and the back-end SOD network, the sampling matrix evolves into specific patterns tailored for SOD. Both simulation and experimental results demonstrate that the proposed method achieves robust and accurate saliency detection even at a low sampling rate of 6.25%. This framework provides a task-oriented paradigm for single-pixel SOD in environments with limited resources.