摘要
目的提出一种有效的大鼠体态识别方法,适用于不同目标分辨率的图像。方法从大鼠体态图像中提取目标的4个旋转、平移、尺度不变量,作为朴素贝叶斯网络的属性变量,将体态分为4类,作为网络的类变量。对网络进行训练,并应用训练好的网络对5组不同分辨率的图像进行识别。结果5组不同分辨率的图像集均取得较高的识别正确率,该方法能有效克服大鼠体态差异等因素带来的目标分辨率不同对识别结果的影响,具有很好的鲁棒性,且运算复杂度低。结论提供了一种大鼠体态识别的算法,该方法具有较强的实用性。
Objective To provide an effective method for rat posture recognition, adaptable to images of different object resolutions. Method A set of effective features invariant to rotation, translation, and scaling were extracted from rat＇ s images and used as naive Bayes classifier＇ s attribute values. And four kinds of rat＇ s postures were categorized as four classes by experienced observers. In our experiment, five groups of image frames with different resolutions were acquired. One was used as training set and the other four as validation set. Result All the five groups with different resolutions reached a high correct recognition rate. The present method with high robust and low computing complexity provided an objective description and classification of rats＇ postures. Conclusion This method has strong applicability and provides an approach for classifying rat postures.
出处
《航天医学与医学工程》
CAS
CSCD
北大核心
2005年第5期370-374,共5页
Space Medicine & Medical Engineering
基金
浙江省国际科技合作项目
国家"十五"重大科技专项项目(2003AAZ3537)的资助
关键词
行为学
体态
朴素贝叶斯分类器
不变量
概率密度估计
ethology
posture
naive Bayes classifier
invariants
probability density estimation
