基于可变形部件模型DPM的目标检测算法采用方向梯度直方图HOG进行特征表示，由于HOG无法处理模糊的边界而且忽略了平滑的特征区域，从而影响了DPM算法的性能。为了提高DPM的性能，提出了一种基于稀疏表示的可变形部件模型目标检测的方法。该方法利用稀疏编码构建一种新的特征描述子来取代原可变形部件所使用的方向梯度直方图，新的特征描述子能够描述物体更多的信息，对图像中的噪声不敏感。实验结果表明，该方法在PASCAL VOC 2012数据集上提高了原可变形部件模型算法的精度。

The object detection method based on the deformable part model (DPM) uses the histogram of oriented gradients (HOG) to describe features. The HOG limits the performance of the DPM, as it cannot deal with noisy edges and ignores the flat areas while focusing on edge areas. In order to improve the performance of the DPM, we propose a DPM object detection method based on sparse representation. Instead of using the HOG, the method uses sparse coding to construct a new feature descriptor. The sparse coding based feature vectors can represent more information of image patches. Experimental results show that the proposed method can improve the precision of the DPM method on the PASCAL VOC 2012 dataset.
相关文章