Image co-segmentation based on pyramid features cross-correlation network

Conclusion In this study, we propose an end-to-end deep learning method to accomplish image co-segmentation pair-wise. The Siamese encoder network is used to extract the high-level features. The core cross-correlation module is based on depth-wise convolution, which models the common semantic information between images from the perspective of feature similarity matching on each channel. And this module can highlight the center position of the high-level features of common objects. A multi-scale feature pyramid is constructed to improve the model’s adaptability for objects of different sizes. We conducted the experiments on several public datasets. The experimental results show that our approach achieves state-of-the-art performance and can well accomplish the image co-segmentation task. Additionally, several groups of ablation experiments are designed to show the segmentation effect under different hyperparameters. The results show a good effect based on the cross-correlation operation of the pyramid features. Please see Appendixes A–C for details.