Dual Attention-Guided Network for Anchor-Free Apple Instance Segmentation in Complex Environments

Abstract

AbstractApple segmentation is an important part of the automatic picking system of apple plantation. However, due to the complexity of apple orchard environments, including light change, branch and leaf occlusion and fruit overlap, the segmentation accuracy of the existing methods is limited, which affects the large-scale application of the automatic picking system. To solve these problems, this paper proposes a new apple instance segmentation method based on a dual attention-guided network. Firstly, the image is preprocessed by the Image Correction Module (ICM) to improve the robustness of the network to the natural environment. Secondly, the Multi-Scale Enhanced Fusion Feature Pyramid Network (MSEF-FPN) is used as the feature extraction module to enhance the ability of image feature extraction, so as to reduce the interference of complex background on apple instance segmentation results without increasing the amount of calculation. Then, a new Dual Attention-Guided Mask (DAGM) branch is added to focus on the pixels of irregular occlusion and overlapping objects, and accurate pixel-level mask segmentation is carried out in the detection rectangular bounding box. Finally, this study carried out instance segmentation experiments on apples with different lighting conditions and different occlusion. The test results show that the model proposed in this paper has excellent detection accuracy, robustness and real-time, and has important reference value for solving the problem of accurate fruit recognition in complex environments.KeywordsApple segmentationComplex environmentsFeature extraction