Single-photon lidar is widely used as an active detection technology with high accuracy and high temporal resolution for 3D high-precision imaging in a variety of scenes. However, weak echo scenarios corresponding to limited signal photon counts and low signal-to-noise ratio scenarios corresponding to high background noise counts pose a great challenge to efficiently and accurately solve the depth. For the single-point ranging scenario of single-photon lidar applied to the above challenging scenarios, this paper proposes a convolutional neural network based on a soft-threshold denoising module and a self-attention mechanism.A convolutional neural network based on a soft-threshold denoising module and a self-attention mechanism is proposed in this paper. The initial feature extraction and data enhancement of the photon sequence histogram data are carried out by the sliding time window module matched with the pulse width of the transmitting laser pulse. And the self-attention mechanism module is introduced to capture the long-range correlation of the photon sequence histogram and to improve the distance solving accuracy and robustness. Then the soft-threshold denoising module is introduced to adaptively generate the threshold value and to filter out the noisy photons, then the echo waveforms of denoised signals are outputted and the depth of the solution is solved. At the same time, this paper uses multi-loss function constraint for network training to focus on the distribution characteristics of the photon sequence histogram and the task demand for a combination of constraints. And we through the ablation experiment to prove its effectiveness. Compared with other histogram techniques, comprehensive experiments on simulated datasets and real datasets show that the proposed model can achieve optimal quantization results, improve the quantization index by at least three times and have better distance resolution performance under different signal-to-noise ratio environments.Based on the comparison of the quantization results of the simulated dataset, the method proposed in this paper is able to identify the signal photon time correlation features and solve the depth with high accuracy and robustness. It can be seen that in the first two signal-to-noise ratio scenarios (2∶10 and 2∶20), removing a small amount of anomalous data, the method proposed in this paper is able to achieve high accuracy and stability with achieving centimeter-level resolution. In the very low signal-to-noise ratio scenario (2∶50), the extractable data features, i.e., the temporal correlation, are affected by a large number of noise photons. The presence of a large number of noise photons gathering anomalously throughout the detection leads to a decrease in the accuracy of the distance solution. Still, the best quantization effect is achieved in the comparison of different methods, which proves the effectiveness and better development prospect of the deep learning method in the weak echo and low signal-to-noise ratio scenarios.For the single-point ranging scenario of single-photon lidar applied to the above challenging scenarios, this paper proposes a convolutional neural network based on a soft-threshold denoising module and a self-attention mechanism. With modules proposed as sliding time window module, self-attention mechanism module, and soft-threshold denoising module, the proposed network achieve high accuracy and stability with achieving centimeter-level resolution in multiple signal-to-noise ratio scenarios in the comparison of different methods.