多尺度区域特征的细粒度分类算法研究

摘要

Intending to reduce the influence of complex background on fine-grained classification, as well as to study the global information and local information of the target objects extracted from the convolu-tional neural network for fine-grained tasks, a fine-grained classification method based on multi-scale re-gion feature was proposed. The method FASTER-RCNN framework was to train three convolution models to locate multi-scale object regions. Then the bounding box constraint and Helen constraint were applied to improve the location accuracy of the detected object. Finally, the extracted multi-scaled region features were combined to train a SVM classifier for fine-grained classification. The proposed method was tested in Caltech-UCSD bird datasets and CompCars vehicle datasets. The results showed that the accuracy of clas-sification in Caltech-UCSD bird datasets was 82.8%. It increased by 7.5 % than the method without multi-scale region features. Compared with part-based RCNN, it increased by 8.9 %. The results showed that the accuracy of classification in CompCars was 93.5%. It increased by 8.3 % than the method with-out multi-scale region features. Compared with GoogleNet, it increased by 2.3 %.%为了避免复杂背景对精细目标分类的影响,充分利用卷积神经网络提取的目标全局和局部信息进行细粒度任务的研究,提出了一种多尺度区域特征的细粒度目标检测与分类算法.该方法先使用FASTER-RCNN框架训练3个尺度区域的卷积模型进行多尺度目标区域定位,对定位的结果进行包围盒约束和海伦约束以优化提高定位的精确度,然后将提取多个尺度区域的特征进行组合,并用支持向量机训练细粒度分类器.在Caltech-UCSD鸟类数据集和CompCars车型数据集上进行实验测试.实验结果表明该算法在 Caltech-UCSD鸟类数据集的分类正确率达到82.8%,比没有使用多尺度区域特征的分类算法提高了7.5%,比基于部件的分类方法提高了8.9%;在CompCars车型数据集的分类正确率达到93.5%,比没有使用多尺度区域特征的分类算法提高了8.3%,比最优的GoogleNet精细目标分类算法提高了2.3%,验证了该算法的有效性.