Dual-path decoder architecture for semantic segmentation of wheat ears

Abstract

In this study, a dual-path decoder segmentation network (DPDS) is presented, which innovatively introduces a dual-path structure into a semantic segmentation network incorporating atrous spatial pyramid pooling (ASPP). A novel loss function, boundary focal loss (BFLoss), is designed specifically for wheat ears segmentation scenarios, which adaptively adjusts weights for different pixel points through the binarization of boundary information, focusing the training on the edges of wheat ears. It is suggested to apply the DPDS network in conjunction with BFLoss to the semantic segmentation of wheat ears. The experimental results demonstrated that BFLoss possesses advantages over commonly used binary cross entropy loss (BCELoss) and focal loss in semantic segmentation. Additionally, the dual-path decoder architecture was proved to reach higher precision than activating only one of the pathways. In comparative experiments with established semantic segmentation networks, the DPDS model achieved the best performance on several evaluation metrics, and attained a balance between precision and recall. Notably, the combination of DPDS and BFLoss achieved a 91.86% F1 score on the wheat ears semantic segmentation test dataset. Therefore, the DPDS model can be effectively applied to semantic segmentation scenarios of crops like wheat, and also provides new insights for the improvement of existing networks. Code is available at https://github.com/awesome-pythoner/dual-path-decoder-segment.