Abstract: | “Composition” determines the vividness of the image and its narrative power. Current research on image aesthetics implicitly considers simple composition rules, but no reliable composition classification and image optimization method explicitly considers composition rules. The existing composition classification models are not suitable for snapshots. We propose a composition classification model based on spatial-invariant convolutional neural networks (RSTN) with translation invariance and rotation invariance. It enhances the generalization of the model for snapshots or skewed images. Ultimately, the accuracy of the RSTN model improved by 3% over the Baseline to 90.8762%, and the rotation consistency improved by 16.015%. Furthermore, we classify images into three categories based on their sensitivity to editing: skew-sensitive, translation-sensitive, and non-space-sensitive. We design a set of composition optimization strategies for each composition that can effectively adjust the composition to beautify the image. |