Compressed photon-counting radar is a combination of photon-counting radar technology and single-pixel imaging technology. It has the advantages of low cost and ultra-high sensitivity. However, the reconstruction of high-resolution imaging cost a large number of measurements and iterative calculations, resulting in a long imaging time. At present, deep learning compression reconstruction network has been proved to be able to avoid iterative computation and achieve rapid compression measurement reconstruction, But the deep learning compression reconstruction network reported in the literature uses traditional image processing database of noise-free pictures or adding Gaussian noise to the pictures to train the network, the network is applied to the actual compressed photon counting radar system, the performance needs to be further verified. A synchronous control measurement module based on FPGA was independently designed, a compressed photon-counting radar system was built, a Monte Carlo simulation of compressed photon-counting radar system to produce training data was proposed, and a deep learning compression reconstruction network DFC-Net was designed for joint optimization of sampling and reconstruction. The experimental results show that: at 10%, 15%, 20%, 25%, and 30% sampling rate, the reconstruction performance of DFC-Net is better than the existing reconstruction network Dr2-Net and the traditional compression reconstruction algorithm TVAL3.