Lightweight Blueberry Fruit Recognition Based on Multi-Scale and Attention Fusion NCBAM

Abstract

Blueberries are widely planted because of their rich nutritional value. Due to the problems of dense adhesion and serious occlusion of blueberries during the growth process, the development of automatic blueberry picking has been seriously hindered. Therefore, using deep learning technology to achieve rapid and accurate positioning of blueberries in the case of dense adhesion and serious occlusion is one of the key technologies to achieve the automatic picking of blueberries. To improve the positioning accuracy, this paper designs a blueberry recognition model based on the improved YOLOv5. Firstly, the blueberry dataset is constructed. On this basis, we design a new attention module, NCBAM, to improve the ability of the backbone network to extract blueberry features. Secondly, the small target detection layer is added to improve the multi-scale recognition ability of blueberries. Finally, the C3Ghost module is introduced into the backbone network, which reduces the number of model parameters while ensuring the accuracy, thereby reducing the complexity of the model to a certain extent. In order to verify the effectiveness of the model, this paper conducts experiments on the self-made blueberry dataset, and the mAP is 83.2%, which is 2.4% higher than the original network. It proves that the proposed method is beneficial to improve the blueberry recognition accuracy of the model.