Combining photon counting technology with single-pixel imaging can achieve highly sensitive and low cost photon counting imaging, but there are problems of long sampling time and long reconstruction time. The compressed sampling and reconstruction network, which is based on deep learning, uses the fully connected layer without the offset and activation function as the measurement matrix, achieves faster and higher quality image reconstruction by learning efficient measurement matrices from the data and avoids the huge amount of calculation caused by traditional iterative algorithms. However, when the fully connected layer is used for block compression sensing of high-resolution images, the reconstructed image will produce block artifact. In response to this problem, overlapping block sampling network (Os_net), nested sampling network (Ns_net), and convolution sampling network (Cs_net) were proposed: to replace fully connected layer sampling. In the design of the reconstructed network, the images were reconstructed by using a linear mapping network. The design experiment shows that Cs_net has the best deblocking effect. After Cs_net binarization is applied to a photon counting single-pixel imaging system, the experiment results show that Cs_net de-blocking effect is significantly better than the traditional algorithm TVAL3, and Cs_net has also achieved good results on the reconstruction quality.