基于改进YOLOv5的柑橘目标识别研究

Research on Oranges Target Recognition Based on Improved YOLOv5

文章针对现有的柑橘目标识别存在准确率不高,以及深度学习模型参数量和浮点计算量大的问题,提出基于YOLOv5算法进行三个方面的改进,一是引入轻量化网络Mobilenetv3、ShufflenetV2、Ghost等对YOLOv5的Backbone模块进行改进,二是针对Neck网络的C3部分融入注意力机制进行改进,三是使用Ghost Conv模块来改进Neck网络的Conv模块。最终改进的算法在模型参数和浮点计算量方面均下降为原来的1/7左右,经过优化参数训练后,该模型在测试集上的mAP@0.5达到了0.957。

Aiming at the problems of low accuracy of existing oranges target recognition,as well as the large amount of parameters and floating-point calculations of deep learning models,this paper proposes to improve the YOLOv5 algorithm in three aspects,one is to introduce lightweight networks Mobilenetv3,Shufflenet V2,Ghost,etc.to improve the Backbone module of YOLOv5,the second is to improve the attention mechanism of the C3 part of the neck network,and the third is to use Ghost Conv module to improve the Conv module of the NECK network.Finally,the improved algorithm reduces the amount of model parameters and floating-point computation to about 1/7 of the original,and after training on optimized parameters,the model reaches 0.957 mAP@0.5 on the test dataset.