期刊论文

Apple Target Detection Based on Improved Faster - RCNN Framework of Deep Learning

中文

机械设计与研究

1006-2343

2019

35

5

24-27

10.13952/j.cnki.jofmdr.2019.0268

湖南省教育厅重点资助项目（15A160）；

本校为其他机构

苹果图像的目标检测是研究苹果采摘机器人的关键技术。本研究以自然光源下的苹果图像为研究对象。首先,针对原始RPN结构的3×3单一小滑动窗口摄取特征信息有限问题,设置不同面积、不同尺寸的9个候选框,同时增加1个滑动窗口,以提高感受野的面积与强度,改进深度学习目标检测框架Faster-RCNN,进而搭建深度学习模型。然后,对图像标注目标的实际位置和类别标签,以苹果的测试集作为输出期望,经过训练得到模型的有效权值。实验结果表明,模型迭代3 000次后,改进模型的目标检测准确率为97.6%,而原始模型为95.3%。另外,改进的模型相对原始模型定位精度更高。

Apple image target detection is the key technology to study apple picking robot. This research takes the apple image under the natural light source as the research object. First, to solve the problem of limited feature information got by 3 ×3 single small sliding window of the original RPN structure, 9 candidate boxes with different area and size are set up, and a sliding window is added to improve the area and intensity of the receptive field, and to improve the detection framework of depth learning target, Faster - RCNN, and then to build th...