Digital image quality evaluation based on multi-scale aesthetic features and graph convolutional neural networks

With the rapid growth of social media and visual content platforms, the aesthetic quality evaluation of digital images has gradually become an important foundation for image content screening and recommendation. However, existing methods have shortcomings in modeling the relationship between local semantics and spatial composition of images, making it difficult to fully reflect the subjective beauty of images. Therefore, this study proposes a dual branch collaborative digital image aesthetic perception model for digital image quality evaluation. This model consists of a semantic guided multi-scale perception network and a composition structure perception graph network. The former enhances feature expression through semantic attention mechanism and multi-scale fusion, while the latter uses graph convolution to model spatial relationships between regions. In the performance test, the Pearson linear correlation coefficient of the proposed model in the high segment was 0.884, with a deviation of only 0.06 and a standard deviation of 0.42. In landscape images, the Kendall rank correlation coefficient of the proposed method was 0.797. Finally, the parameter count of this method was 8.6 million, the number of floating-point operations was 3.02 billion, and the inference frame rate was 54.2 FPS, achieving a balance between performance and efficiency. The experiment shows that the proposed model performs outstandingly in terms of rating consistency, semantic perception, and structural modeling. The study attempts to offer a more accurate and comprehensive automated evaluation solution for applications related to image aesthetic evaluation, and to assist in the development of intelligent image understanding and quality control.