在自然环境背景下使用改进R-FCN对香蕉串识别的方法
Improved R-FCN for Banana Bunch Identification in Natural Environment Context
冯嘉倩 1曹洪 1王金权1
作者信息
- 1. 广州现代信息工程职业技术学院/信息工程学院 广东 广州 510663
- 折叠
摘要
香蕉串的检测与识别对智能化果实采集至关重要.然而,果园风景复杂、香蕉和树叶等背景具有相似性使得自动、快速、准确识别香蕉串成为智能化果实采集的挑战.为此提出一种改进的R-FCN模型,用于检测复杂背景下的香蕉串.从香蕉种植基地收集了大量图像,并通过旋转和颜色转换技术进行增强.训练时,这些增强图像与原始图像共同使用.采用经剪裁的ResNet网络和像素加权损失函数,提高了检测速度和分割效果.测试结果表明,该方法在AP50和AP70时的准确率分别达到93.1%和90.4%,检测时间缩短至0.187s,相较于其他图像检测算法,速度更快、精度更高.
Abstract
The detection and identification of banana bunches are crucial for intelligent fruit collection.However,due to the complex orchard landscape and the similarity of banana and leaf backgrounds,the automatic,fast and accurate identification of banana bunches become a challenge for intelligent fruit collection.Therefore,an improved R-FCN model is proposed to detect banana bunches in complex backgrounds.A large number of images are collected from the banana plantation base and enhanced by rotation and color conversion techniques.During the training,these enhanced images are used concomitantly with the original images.The tailored ResNet network and pixel-weighted loss function are used to improve the detection speed and segmentation effect.The test results show that the accuracy of this method at AP50 and AP70 reaches 93.1%and 90.4%,respectively,and the detection time is shortened to 0.187s,which is faster and more accurate than other image detection algorithms.
关键词
自然环境/香蕉串识别/R-FCN/区域全卷积神经网络Key words
natural environment/banana bunch identification/R-FCN/regional fully convolutional neural network引用本文复制引用
出版年
2024
NETL
NSTL
万方数据