Air-gapped computers are separated from the Internet and unable to exfiltrate data via storage mediums such as flash memory. Recent research indicates that physical transmission channels can be established to exfiltrate data from an air-gapped computer using acoustical, optical, thermodynamic, and other physical methods. However, the transmission speed of these methods is limited. The rapid development of smartphone cameras has made it easy to record high-quality videos of screens. Inspired by it, we propose a novel deep learning-based covert image exfiltration method, which enables high-speed covert transmission of grayscale images from an air-gapped computer through the smartphone camera. We utilize the human eye’s persistence of vision effect, embed distinct grayscale images into individual frames of natural color images, and subsequently capture the screen using smart devices, such as smartphones, thereby achieving the covert transmission of secret images. On a screen with a refresh rate of 60 Hz, by selecting a rectangular region, our method can covertly transmit up to 60 grayscale secret images, each of the same size, per second. Additionally, an edge mask based on erosion operation is proposed to enhance the robustness and imperceptibility of our method. Experimental results show its effectiveness under various screen and recording conditions.