关键词: 草莓图像, 实例分割, 改进Mask R—CNN, CA注意力机制, 轻量化网络

Abstract: In response to the complexity of the orchard picking environment, the difficulty of accurate segmentation between strawberries and the surrounding environment, and the inability of the existing model processing speed to realize fast segmentation, a segmentation algorithm of lightweight strawberry instance based on Mask R—CNN is proposed. On the basis of the original Mask R—CNN algorithm, the MobileNetV3 network is used to replace the original ResNet101 backbone network, the algorithm is lightweight, and the channel attention mechanism in the original MobileNetV3 residual structure is replaced by the collaborative attention mechanism module, which is combined with the feature pyramid network architecture to perform the feature extraction, and the strawberry individuals are realized. The precise and fast localization segmentation of individual strawberries is achieved. Finally, comparison experiments are performed on the self-labeled dataset. The experimental results show that the proposed improved Mask R—CNN algorithm improves border mAP and mask mAP by 1.75% and 4.05% respectively， and the detection speed by 20.09 frames/s compared with the original Mask R—CNN model, which reduces the dependence of the model on hardware storage space and arithmetic power.

Key words: strawberry image, instance segmentation, improved Mask R—CNN, CA attention mechanism, lightweight network