Image Cropping Based on Aesthetic and Saliency Perception Regression

Key Points

• This research aims to develop an interpretable image cropping method that integrates aesthetic and saliency perception for improved outcomes.
• Constructed a feature extraction network to generate aesthetic and visual saliency maps.
• Integrated spatial loss to adjust aesthetic and saliency features for key compositional elements.
• Designed an anchor-guided regressor for predicting crop boxes from coarse to fine resolution.
• Achieved an average IoU improvement of approximately 1.72% on the FLMS dataset.
• Reduced mBDE by about 17.05% on the FLMS dataset, outperforming existing methods.