Bounding Box Regression Based Image Composition Recommendation

Image composition recommendation aims to find the most aesthetically valuable crop in an image,which can assist the photographer to take beautiful,elegant,and coordinated photos.However,most of the previous methods based on neural network do not generate the composition boxes directly because it is difficult to accurately and completely mark all excellent composition bounding boxes.Instead,these methods first enumerate some pre-made boxes,regress their score and return the box with highest score,which will have a negative impact on the accuracy of regression results and algorithm efficiency.By proposing an end-to-end neural network model consisting of a feature extraction module,a bounding box regression module,and a score regression module,which can regress the composition bounding box,and designing the corresponding data construction method,training method and loss function,the above difficulties have been successfully overcome.In this paper,two public datasets,FCDB and FLMS,are selected as the test set.Compared with the existing methods,this method achieves the best IoU and Disp.