基于多特征融合的自适应权重目标分类方法  

Target classification with adaptive weights based on multi-feature fusion

作　　者：王立鹏[1] 张智[1] 苏丽[1] 聂文昌 WANG Lipeng;ZHANG Zhi;SU Li;NIE Wenchang(College of Automation,Harbin Engineering University,Harbin 150001,China)

机构地区：[1]哈尔滨工程大学自动化学院,黑龙江哈尔滨150001

出　　处：《华中科技大学学报（自然科学版）》2020年第9期38-43,共6页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：国家自然科学基金资助项目(61803116);中央高校基本科研业务费专项资金资助项目(3072020CF0410)。

摘　　要：提出一种自适应权重的融合卷积特征和方向梯度直方图(HOG)特征的目标分类方法,实现快速精准分类目的.首先,利用ResNet网络框架提取图像卷积特征,增加OpenCV接口以提取图像HOG特征,对HOG特征图扩维处理至与卷积特征同维;然后,在ResNet网络框架中嵌入SENet模块,计算卷积特征和HOG特征的权重向量,将卷积特征、HOG特征及相应的权重向量加以变权值叠加,实现多特征的自适应同步融合,以此构建二分类网络模块;其次,将二分类网络模块嵌入Faster Rcnn网络,构成Faster Rcnn-HOG新型网络,通过基于变阈值的粗检测策略和先验知识的区域关注策略得到图像中目标预处理检测框,利用二分类网络模块精确判定,实现目标分类.将Faster Rcnn-HOG与传统Faster Rcnn网络及另一特征融合网络Net-BB-HOG进行对比试验,三种方法在目标大类识别方面性能基本相当,但是FasterRcnn-HOG在目标小类识别方面效果更佳,证明了提出的多特征融合自适应目标分类方法的有效性和正确性.To solve precision forming problem of the antenna panels with large non-developable double,a target classification method with adaptive weights was proposed on the basis of fusion of convolution feature and histogram of oriented gradient(HOG)feature,which was utilized to classify the targets quickly and precisely.First of all,convolution feature was extracted through the ResNet framework,in which the OpenCV interface was increased to acquire the HOG feature of the images.The dimensions of HOG feature were enlarged to maintain the same dimensions as the convolution feature.Second,SENet module was imbedded into the ResNet framework so that the weight vectors of the convolution feature and HOG feature were calculated.The features of the images were adaptively and synchronously fused based on the convolution feature,HOG feature,and the weight vectors.An innovation binary network was established based on the multi-feature fusion.Third,the binary network was imbedded into the Faster Rcnn network to establish Faster Rcnn-HOG,in which the pre-processing detection frames of the image was acquired through the strategies of coarse detection of variable threshold and focus area of prior knowledge.Then the pre-processing detection frames was precisely judged by the proposed binary network to realize the target classification.The comparative experiments among faster Rcnn-HOG,the traditional Faster Rcnn,and another feature fusion network Net-BB-HOG were conducted.The results verify that the effect of the three methods is similar in the target classification of large categories.However,faster Rcnn-HOG is more effective in identifying small categories of the targets.The validity and correctness of the proposed method is proved.

关 键 词：目标分类 深度学习 卷积神经网络 多特征融合 自适应权重 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]