多尺度特征融合与极限学习机的玉米种子检测

Detection of maize seeds based on multi-scale feature fusion and extreme learning machine

目的玉米种子等农作物检测是农业信息化领域中一个关键问题,为了能够快速和准确地实现对其检测,提出基于多尺度特征融合与极限学习机的玉米种子无损检测算法。方法首先对种子特征的描述采用局部特征和全局特征相结合的特点,局部特征采用多尺度方向梯度直方图特征,而在全局特征上则提取HSV颜色模型特征。其次,针对传统的BP神经网络以及SVM等存在训练时间长、检测速度慢的不足,采用极限学习机作为其检测算法。此外,为了避免样本在训练时带来的过多时间消耗,该检测模型上采用了并行训练算法。再次,针对原始图像分辨率过高问题所带来的检测时间以及内存消耗较大的问题,采用一种基于局部均值的图像缩小算法。最后,针对该文采用的滑动窗口扫描可能带来的同一对象多窗口重叠的问题,提出了一种基于模糊聚类的局部窗口融合算法。结果仿真结果表明,提出的方法可实现对玉米种子的准确检测,对检测样本的测试精度达到97.66%,同时误差控制在0.1%。结论相比传统的方法,提出的方法在检测速度、检测准确率上均有所改善,无需严格的硬件设备要求并且对玉米种子检测时不会产生任何损伤。

Objective The detectionof maize seeds and other crops is a key problem in the field of agricultural infonnatiza- tion. In this paper, to enhance detection and improve accuracy rate, we present a detection algorithm for maize seed on the basis of extreme learning machine （ELM） and multi-scale feature fusion. Method In the first part of this paper, the feature of maize seed is described as the combination of local features and global features. Local features can be described as a multi-scale histogram of oriented gradient, and global features can be described as the color feature of HSV. In the second part, ELM will be used as the detection algorithm against long training period and slow detection speed, which are the char- acteristics of traditional BP neural network and SVM. Furthermore, the detection model uses the parallel algorithm to signif- icantly decrease the time used in the training of each classifier. Furthermore, the high resolution of the original image can cause long detection periods and consume a large amount of memory. To address this problem, we propose a local means ofan image compression algorithm. Finally, considering that the sliding windows of centralized scanning can produce a prob- lem in creating multiple overlapping windows capturing the same object, we propose a local window fusion algorithm based on fuzzy clustering to address this problem. Result The simulation results show that the method proposed is able to accu- rately detect maize seeds. The accuracy of detection of maize seeds can reach 97. 66% with an error of less than 0. 1%. Conclusion Compared with traditional methods, the no damage method proposed in this paper can improve the speed and accuracy of detection and has no strict hardware requirements.