融合残差网络与注意力机制的草莓检测

Strawberry detection combining residual network with attention mechanism

针对草莓果实因受到自然光光照、枝叶遮挡、果实间存在遮挡等因素,较难实现成熟草莓果实识别的现状,提出融合深度残差网络与注意力机制的成熟草莓目标检测算法。引用信息表达能力更强的深度残差网络Resnet50对SSD目标检测算法模型基础骨干网络进行替换,对经过残差网络结构和新增卷积特征提取层得到信息特征提取图进行通道和空间方向的注意力机制方法处理,建立能准确实现成熟草莓目标检测的RC-SSD目标检测模型。试验结果表明,本文的RC-SSD算法模型对比Faster R-CNN、YOLOv3、SSD-VGG模型拥有较少的参数量,平均精度均值mAP分别提升46.05%、10.16%、5.77%,其中成熟草莓的识别精度达到99.04%。对比轻量化网络结构模型SSD-Mobilenetv2,RC-SSD算法模型在FPS相对于轻量化网络模型降低25帧的情况下,精度提升20.20%,FPS在GPU运行设备上达到86帧。

In view of the current situation that it is difficult to recognize ripe strawberry fruit due to the factors such as natural light illumination,branch and leaf shading,and inter-fruit shading,this paper proposes a ripe strawberry target detection algorithm that combines deep residual network and attention mechanism.In this paper,the deep residual network Resnet50,which had stronger information expression capability,was invoked to replace the backbone network underlying the SSD target detection algorithm model,and the attention mechanism method of channel and spatial direction was processed to obtain the information feature extraction map after the residual network structure and the new convolutional feature extraction layer,and the RC-SSD target detection model that could accurately implement the mature strawberry target detection was established.The experimental results showed that the RC-SSD algorithm model in this paper had less number of parameters than the models Faster R-CNN,YOLOv3 and SSD-VGG models,and the average accuracy mean mAP was improved by 46.05%,10.16%and 5.77%,respectively,in which the recognition accuracy of mature strawberry reached 99.04%,and compared with the lightweight network structure model SSD-Mobilenetv2,the RC-SSD algorithm model improved the accuracy by 20.20%with a 25 fps reduction in FPS relative to the lightweight network model,and the FPS reached 86 fps on the GPU running device.