Abstract: Aiming at the problems of unclear intrinsic connection between aesthetics and cropping and lack of compositional integrity in existing image cropping methods, this paper proposes an interpretable aesthetic and saliency-aware image cropping method. Firstly, a feature extraction network is constructed to generate the aesthetic region distribution map and the visual saliency feature map respectively; secondly, the two types of features are fused and adjusted through the spatial receptive loss to locate the key compositional elements; finally, an anchor-guided regressor is designed to implement the cropping box prediction from coarse-grained to fine-grained. Experiments on the FCDB and FLMS datasets show that while maintaining the integrity of the main subject, the accuracy of image cropping of this method is superior to that of the existing mainstream methods. On the FLMS dataset, the IoU is increased by approximately 1.72% on average, and the mBDE is improved by approximately 17.05% on average.