In the field of computer graphics, achieving real-time, physically-based global illumination is a highly challenging task. This paper presents an efficient method for real-time global illumination based on light probes. It integrates three key techniques: adaptive screen-space probe placement, efficient importance sampling, and neural network-based image reconstruction. The adaptive screen-space probe placement strategy breaks away from conventional uniform 3D placement, which distributes probes according to scene characteristics, improving the utilization efficiency of light probes. As a result, it reduces the overall number of probes while maintaining a high-quality illumination approximation, thus saving computational resources. Efficient importance sampling is based on Monte Carlo integration. Considering the incoming light radiance and the normal direction, it samples more in regions crucial for the final rendering. This method reduces the number of samples and enhances the accuracy of illumination calculations. Due to the limited number of light probes and low sampling rate, the rendering may suffer from noise and aliasing. To address this, a lightweight neural network is designed for image reconstruction. The neural network generates high-quality global illumination images from low-quality probe data, improving the visual effect of rendering. Experimental results on four well known test scenes show that our method outperforms traditional methods in terms of rendering quality and running efficiency, providing a more effective solution for real-time global illumination rendering.