Due to its unique historical evolution, the painting techniques of the Lingnan region in modern times have been influenced to a certain extent by Western culture. This paper obtained approximately 3,000 images of modern Lingnan landscape paintings and Western landscape paintings through on-site interviews and online collection. The images were rotated, cropped, and scaled to enhance the data representation of the paintings and complete the preprocessing of the research data. In terms of painting image classification, the optimal feature subset was selected based on the CGO optimization algorithm, and the cross-task feature fusion module was used to fuse painting image features. This enabled the construction of a multi-faceted artistic painting classification model, achieving the classification task of different painting types under a unified framework. In terms of painting image emotional classification, self-learning and knowledge transfer techniques based on sparse autoencoders were introduced as methods for painting image emotional semantic analysis and unsupervised feature learning. Combining the image features of modern and contemporary Lingnan landscape paintings and Western landscape paintings, we propose a painting image emotional classification system framework comprising three major modules: source domain local feature learning, target domain global feature extraction, and image emotional classification. This framework is used to construct a painting emotional classification model. The designed painting emotion classification model not only demonstrates emotion classification accuracy significantly higher than similar models (>0.730) but also achieves a classification performance standard deviation <0.010 after oversampling strategies, demonstrating excellent robustness. This provides a robust and effective technical foundation for analyzing the artistic ambiance of painting images.