The revitalization of local cultural heritage provides impetus for regional cultural development. This paper utilizes a cyclic threshold selection algorithm and a convolutional neural network (CNN) classification model to enhance the accuracy of 3D modeling and video animation scanning recognition of local cultural heritage educational books using augmented reality (AR) technology. The algorithm calculates an adaptive threshold for images, segments binary images, and combines the cosine loss function of image feature vectors to optimize the model’s ability to correctly classify and differentiate between frames. AR scanning and recognition of local cultural heritage educational books are implemented on both the Vuforia and Unity platforms to enhance the immersive reading experience. Research findings indicate that when the number of pooling layers and convolutional layers are 4 and 3, respectively, the model achieves the lowest loss function value and the highest scores across all seven evaluation metrics. The average values for 4pt-Homography RMSE, PSNR, and SSIM are 0.5524, 30.7333, and 0.9365, respectively, with registration classification performance superior to that of the comparison model. The average scores of the experimental group students were significantly higher than those of the control group at the 0.01 level, and their satisfaction with the AR-based auxiliary teaching method reached 100%.
