Interactive Segmentation Method Based on Multi-Network Hierarchical Training

Existing interactive segmentation methods generally adopt iterative methods to train the network. These methods randomly select some clicks as the initial clicks at the beginning of the training. Since these clicks are randomly selected, the corresponding situations of these clicks cannot be trained during the training, resulting in the segmentation results corresponding to these clicks not being optimal, and further leading to the results corresponding to the initial several user clicks in practical applications not being optimal. In response to the above problems in the iterative training method, this paper proposes an interactive segmentation model based on multi-network hierarchical training. The training strategy of layer-by-layer training is used to optimize the selection of click samples, multiple networks are utilized to obtain more complete and fine-grained segmentation results, and a complex backbone network and a lightweight network fusion module are designed to reduce the algorithm complexity. The experimental results on the Berkeley, DAVIS, and SBD datasets show that compared with methods such as RITM and GrabCut, our method has higher segmentation accuracy.