Coded aperture radiation imaging technology has been extensively applied in nuclear security, decommissioning of nuclear facilities, radioactive decontamination, and special nuclear material detection. However, in complex radioactive environment, common decoding algorithms have poor ability to suppress noise. In this paper, an reconstruction algorithm based on convolutional neural networks (CNN) is proposed. Additionally, the Geant4 software is utilized to simulate the process of encoded aperture imaging. Meanwhile numerical cut method was used to suppress noise. The reconstruction results show that the average CNR of orphan source reconstructed by the CNN method is 15.2, while that of the cross-correlation decoding algorithm and the dual cross-correlation decoding algorithm are 13.1 and 5.8, respectively. Moreover, after using numerical cut method, the CNN decoding algorithm can obtain ideal image reconstruction results for 94.2% of the radioactive sources in the field of view (FOV). When there are 5 radioactive sources in the FOV, the image reconstructed by cross-correlation decoding algorithm contains artifacts, while the CNN algorithm can accurately distinguish the source. Therefore, under complex conditions involving multiple radioactive sources, the convolutional neural network algorithm has stronger adaptability than the cross-correlation decoding algorithm and can perform more complex and accurate positioning.: Coded aperture radiation imaging technology has been extensively applied in nuclear security, decommissioning of nuclear facilities, radioactive decontamination, and special nuclear material detection. However, in complex radioactive environment, common decoding algorithms have poor ability to suppress noise. In this paper, an reconstruction algorithm based on convolutional neural networks (CNN) is proposed. Additionally, the Geant4 software is utilized to simulate the process of encoded aperture imaging. Meanwhile numerical cut method was used to suppress noise. The reconstruction results show that the average CNR of orphan source reconstructed by the CNN method is 15.2, while that of the cross-correlation decoding algorithm and the dual cross-correlation decoding algorithm are 13.1 and 5.8, respectively. Moreover, after using numerical cut method, the CNN decoding algorithm can obtain ideal image reconstruction results for 94.2% of the radioactive sources in the field of view (FOV). When there are 5 radioactive sources in the FOV, the image reconstructed by cross-correlation decoding algorithm contains artifacts, while the CNN algorithm can accurately distinguish the source. Therefore, under complex conditions involving multiple radioactive sources, the convolutional neural network algorithm has stronger adaptability than the cross-correlation decoding algorithm and can perform more complex and accurate positioning.

Coded Aperture Imaging (CAI); Monte Carlo simulation; Cross-Correlation Decoding; Convolutional neural network (CNN)