基于注意力机制与改进YOLO的温室番茄快速识别

为了实现复杂环境下农业机器人对番茄果实的快速准确识别,提出了一种基于注意力机制与改进YOLO v5s的温室番茄目标快速检测方法。根据YOLO v5s模型小、速度快等特点,在骨干网络中加入卷积注意力模块(CBAM),通过串联空间注意力模块和通道注意力模块,对绿色番茄目标特征给予更多的关注,提高识别精度,解决绿色番茄在相似颜色背景中难识别问题;通过将CIoU Loss替换GIoU Loss作为算法的损失函数,在提高边界框回归速率的同时提高果实目标定位精度。试验结果表明,CB-YOLO网络模型对温室环境下红色番茄检测精度、绿色番茄检测精度、平均精度均值分别为99.88%、99.18%和99.53%,果实检测精度和平均精度均值高于Faster R-CNN模型、YOLO v4-tiny模型和YOLO v5模型。将CB-YOLO模型部署到安卓手机端,通过不同型号手机测试,验证了模型在移动终端设备上运行的稳定性,可为设施环境下基于移动边缘计算的机器人目标识别及采收作业提供技术支持。

Fast Recognition of Greenhouse Tomato Targets Based on Attention Mechanism and Improved YOLO

In order to realize the rapid and accurate recognition of greenhouse tomato fruit by agricultural picking robot in the complicated environment of greenhouse, a fast target detection method for greenhouse tomato fruit based on attention mechanism and improved YOLO v5s was proposed. According to the characteristics of small size and fast speed of YOLO v5s(You only look once v5s) model, the convolutional block attention module (CBAM) was added into the backbone network. By concatenating spatial attention module and channel attention module, the problem of color similarity between green tomato fruit and its background was solved. More attention was paid to the target features of green tomato fruit to improve the recognition accuracy. Replacing GIoU Loss with CIoU Loss as the new loss function of the algorithm contributed to improve the positioning accuracy while improving the bounding box regression rate. The test results showed that the recognition accuracy of the CB-YOLO network model for red tomato fruit detecting precision and green tomato fruit detecting precision and mean average precision in greenhouse environment was 99.88%, 98.18% and 99.53%, respectively. Compared with Faster R-CNN network model, YOLO v4-tiny network model and YOLO v5 network model, the detection accuracy and the mean average precision were improved. The CB-YOLO model was deployed to Android system of mobile phones after being tested by different mobile phones, which verified the stability of the performance detection of the deployment model under actual working condition. It will provide technical support for target detection and harvesting based on robotic mobile edge computing in facility environments.