融合多源图像信息的果实识别方法  

Fruit Recognition Algorithm Based on Multi-source Images Fusion

作　　者：冯娟[1,2] 曾立华[3] 刘刚[2] 司永胜[1] 

机构地区：[1]河北农业大学信息科学与技术学院,保定071001 [2]中国农业大学现代精细农业系统集成研究教育部重点实验室,北京100083 [3]河北农业大学机电工程学院,保定071001

出　　处：《农业机械学报》2014年第2期73-80,共8页Transactions of the Chinese Society for Agricultural Machinery

基　　金：国家自然科学基金资助项目(31371532;31371532-B);河北省科技计划自筹经费项目(13237210);保定市科学技术研究与发展计划资助项目(12ZG011)

摘　　要：光线变化与目标重叠是影响自然环境中果实正确识别的重要原因。为了降低两者的影响，研究了融合多源图像信息的果实识别方法。在图像配准的基础上，优选了日分量图与幅度图像作为待融合的源图像；由模糊推理系统（隶属度函数和模糊规则）决定权重，采用加权平均策略实现图像的像素级融合；根据融合图像中果实区域的分布规律，设计了一种基于直方图的首阈检测法以获得最佳的果实分割效果；利用深度图像的统计特性，设计了一种逐层分割图像的方法以解决重叠果实的分离问题。实验结果表明：多源融合图像用于果实识别与定位比单一图像具有更好的准确性与鲁棒性，对重叠果实的正确识别率在83．67％～94．22％之间。Light changing and targets overlapped each other were main reasons of affecting fruit recognition accuracy under natural condition. In order to reduce both effects, a kind of fruit recognition algorithm based on multi-source images fusion was studied. On the basis of image registration, H component image and amplitude image were selected as source images for fusion. The use of fuzzy logic in pixel level fusion was related to the weighted averaging approach where the weights were determined by using the fuzzy inference system （membership function and fuzzy rules）. According to the law of fruit area distribution in the fused image, an head threshold detection algorithm based on histogram was presented,so as to get the best fruit segmentation results. According to statisitical properties of range image, a solution for overlapped fruits recognition using layer segmentation algorithm was designed. The experimental results showed that information on multi-source image fusion was used for fruit recognition and location more accurately and robustly than single image, overlapped fruit recognition rate was from 83.67% to 94.22%.

关 键 词：果实识别 多源图像 模糊逻辑 图像融合 图像分割 

分 类 号：TP242.62[自动化与计算机技术—检测技术与自动化装置]