机器人采摘苹果果实的K-means和GA-RBF-LMS神经网络识别

为进一步提升苹果果实的识别精度和速度,从而提高苹果采摘机器人的采摘效率。提出一种基于K-means聚类分割和基于遗传算法(genetic algorithm,GA)、最小均方差算法(least mean square,LMS)优化的径向基(radial basis function,RBF)神经网络相结合的苹果识别方法。首先将采集到的苹果图像在Lab颜色空间下利用K-means聚类算法对其进行分割,分别提取分割图像的RGB、HSI颜色特征分量和圆方差、致密度、周长平方面积比、Hu不变矩形状特征分量。将提取的16个特征作为神经网络的输入,对RBF神经网络进行训练,以得到苹果果实的识别模型。针对RBF神经网络学习率低、过拟合等不足,引入遗传算法对RBF隐层神经元个数和连接权值进行优化,采取二者混合编码同时进化的优化方式,最后再利用LMS对连接权值进一步学习,建立新的神经网络优化模型(GA-RBF-LMS),以提高神经网络的运行效率和识别精度。为了获得更精确的网络模型,在训练过程中,苹果果实连同树枝、树叶一块训练;得到的模型在识别过程中,可一定程度上避免枝叶遮挡对果实识别的影响。为了更好地验证新方法,分别与传统的BP(back propagation)和RBF神经网络、GA-RBF优化模型比较,结果表明,该文算法对于遮挡、重叠果实的识别率达95.38%、96.17%,总体识别率达96.95%;从训练时间看,该文算法虽耗时较长,用150个样本进行训练平均耗时4.412 s,但训练成功率可达100%,且节省了人工尝试构造网络结构造成的时间浪费;从识别时间看,该文算法识别179个苹果的时间为1.75 s。可见GA-RBF-LMS网络模型在运行效率和识别精度较优。研究结果为苹果采摘机器人快速、精准识别果实提供参考。

Apple recognition based on K-means and GA-RBF-LMS neural network applicated in harvesting robot

Abstract: In order to improve the recognition precision and speed for apple, and further improve the harvesting efficiency of apple harvesting robot, an apple recognition method based on combining K-means clustering segmentation with genetic radial basis function (RBF) neural network is proposed. Firstly, the captured apple image is transformed into L*a*b* color space, and then under this color space, the K-means clustering algorithm is used to segment the apple image. The color feature components and shape components of segmented image are extracted respectively. The color features include R, G, B, H, S and I, a total of 6 feature components; and the shape features include circular variance, density, ratio of perimeter square to area, and 7 Hu invariant moments, a total of 10 shape components. These extracted 16 features are used as the inputs of neural network to train RBF neural network, and get the apple recognition model. Due to some inherent defects the RBF neural network has, such as low learning rate, easily causing over fitting phenomenon, genetic algorithm (GA) is introduced to optimize the connection weights and the number of hidden layer neurons. In this study, a new optimization way is adopted, that is, the hybrid encoding of the number of hidden layer neurons and connection weights is carried out simultaneously. This moment, the learning of weights is not completed, and the least mean square (LMS) is used to further learn the connection weights. Finally, an optimized neural network model (GA-RBF-LMS) is established, which is to improve the operating efficiency and recognition precision. In the experiments, there are 150 images captured, and they have 229 apples; among them 50 images are selected as training samples, and the rest as testing samples. Every image for training sample has only one apple, so the testing samples have 179 apples. In order to get the precise model, fruits of apple are together with branches and leaves for training during the training process, which avoids the influence of branches or leaves shade on the recognition to some extent. So the training samples have 50 apples, 50 branches and 50 leaves, which are a total of 150 training samples, and the outputs of neural network include 3 classes. In order to compare with the traditional back propagation (BP) and RBF neural network, and GA-RBF algorithm, a series of experiments are carried out. After repeated trainings of 50 times, the results show that the successful training rate of the GA-RBF-LMS is the highest, which can reach 100% and get the minimum training error; but its running time is the longest, because the 2 optimizations of genetic algorithm and LMS are at the expense of the time. The recognition rates of the fruits with different growth postures, such as fruit without obscuration, overlapping fruit and covered fruit, are calculated respectively. After repeated experiments of 50 times, the results show that these 4 recognition models can achieve very good effect for recognizing the fruit without obscuration. For covered fruit and overlapping fruit, the recognition rate of GA-RBF-LMS is the highest, which can reach 95.38% and 96.17%, respectively. Looking from the overall, the recognition rate reaches 96.95%, recognizing 179 apples consumes 1.75 s, and the sum of square of error is the smallest. From the training time, the GA-RBF-LMS algorithm is the longest, whose average training time is 4.412 s for 150 training samples, but the training success rate can reach 100%, which saves the time wasted in human trying to construct the network structure. All of these illustrate that the GA-RBF-LMS neural network model has the higher operating efficiency and recognition precision, and it can be applied in target recognition for apple harvesting robot.