In this paper, we construct a dataset of modern Lingnan landscape paintings and Western landscape paintings in cross-cultural perspective from multiple aspects, and ensure the dataset usability through preprocessing operations. Aiming at the problem that the traditional CycleGAN model cannot retain more contextual detail information, three loss functions are introduced on the basis of the loss functions of the generative network and the discriminative network, and a detailed training process is also given, aiming at retaining the contextual features of the artworks. Subsequently, it is inputted into the convolutional neural network based on hybrid attention mechanism to extract the contextual features of each frame of modern Lingnan landscape painting and western landscape painting images, and then it is put into the bidirectional GRU model to learn the contextual information, and finally realizes the contextual classification of the images of modern Lingnan landscape painting and western landscape painting, and thus designs a contextual classification model based on CNN-BiGRU-At. Compared with the CycleGAN model, the improved CycleGAN model has a higher priority in image digitization conversion, and its values are classified as 0.759~0.941. In addition, the accuracy distribution of the context classification of digital images of modern Lingnan landscape paintings of the CNN-BiGRU-At model ranges from 0.5 to 0.73, and the accuracy distribution of digital images of western landscape paintings of the context classification is 0.7~0.9, which is 0.5~0.73. 0.7~0.9, clearly perceived modern Lingnan landscape paintings and western landscape paintings to create differences in mood, western landscape paintings in general to simple, abstract mood is mainly gradual to attract people to think deeply, Lingnan landscape paintings are complexity and richness, visual elements have rocks, fishing village, smoke and rain, boat trip, etc., the overall mood for the rich and concrete mainly.
