摘要
可形变部件模型(DPM)在目标检测已取得较好的效果,但因为现有的目标检测数据集中鸟类样本数量过少,分布不均衡,而且采用的HOG特征无法较好的表征鸟类目标,造成自然场景鸟类检测的准确率很低。针对这个问题,本研究首先对Image Net数据集上的鸟类样本进行筛选和数据分析,选取自然场景中鸟类样本1 500个,生成对应的标注文件,建立了自然场景鸟类数据集;并提出一种结合多通道特征(Aggregate Channel Features,ACF)和DPM的自然场景鸟类检测方法,算法从自然场景鸟类数据集中的训练样本中,提取ACF特征,再通过Latent SVM训练得到ACF-DPM模型;进一步研究了模型组件和部件个数对于鸟类检测效果的影响。实验结果表明:本文算法在复杂的自然场景中,能够有效地进行鸟类检测,整体精度优于传统的DPM算法。
Deformable Part Model（ DPM） has achieved good results in target detection,but because of lacking bird sample and unbalanced training set,and single HOG feature cannot characterize bird target exactly,so the algorithm cannot meet high bird detecting accuracy in nature scenes. According to these problems,firstly,this paper filter and analyze the bird samples from Image Net dataset,then choose 1 500 bird samples from natural scenes,generating the corresponding annotation files,then set up a dedicated bird data set. A new kind of bird detection in nature scene method which combines with DPM and Aggregate Channel Features is proposed. The algorithm extracts ACF（ Aggregate Channel Features,ACF） from training samples of the dedicated bird dataset,and gets ACF-DPM model through Latent SVM training. The effects of component and part numbers on bird detection are further researched. Experimental results show that even in a complex nature scene,bird detection can be done effectively with this algorithm,the overall accuracy of which is better than traditional DPM algorithm.
作者
张逸扬
储珺
王璐
ZHANG Yu-yang;CHU Jun;WANG Lu(School of Software, Nanchang Hangkong University, Nanchang 330063, Chin)
出处
《南昌航空大学学报（自然科学版）》
CAS
2018年第1期76-82,89,共8页
Journal of Nanchang Hangkong University(Natural Sciences)
基金
国家自然科学基金(61663031
61663036)
江西省重点研发计划项目(20161BBE50085)
关键词
鸟类检测
自然场景
DPM
ACF特征
bird detection
nature scenes
deformable part model
aggregate channel features
