Three-dimensional combustion diagnosis based on tomography technology requires the use of flame images in multiple directions for three-dimensional reconstruction. The accuracy of lens angle calibration directly affects the qua-lity and precision of reconstruction. In order to reduce the error of angle calibration, a Convolutional Neural Network (CNN) lens angle calibration method based on cross calibration block is proposed. A new cross calibration block was designed. Compared with the traditional calibration block, it has more complex spatial structure characteristics, which can enhance the geometric information difference of the image during rotation, help CNN to extract angle features more accurately, and reduce the angle label error in the training set. A CNN based on the Residual Neural Network (ResNet) architecture was built for angle prediction, and CNN training was implemented based on the open-source framework PyTorch to avoid artificial feature design. Experiments were conducted to verify the application effect of the CNN lens angle calibration method based on the cross calibration block. The results show that when the traditional triangular prism calibration block is used for lens angle calibration, the angle label error is large, resulting in low accuracy of lens angle calibration; when the cross calibration block is used for lens angle calibration, the loss function converges faster during model training and the accuracy of lens angle calibration is higher. The convolutional neural network lens angle calibration method based on the cross calibration block shows higher robustness and stronger generalization ability, providing technical support for improving the accuracy of combustion image reconstruction.