基于深度卷积特征与LLC编码的现勘图像分类

Crime scene investigation image classification based on deep convolution features and LLC coding

为了改善低层特征对图像内容描述不够精确而导致现勘图像分类准确率低的问题,提出一种利用深度学习特征的改进局部约束线性编码(local-constrained linear coding,LLC)算法。采用滑动窗口法提取图像密集卷积神经网络(convolutional neural networks,CNN)特征;利用近似LLC算法对提取的密集CNN特征进行快速编码和最大池化,并采用多尺度空间金字塔匹配产生包含空间位置信息的稀疏编码特征。最后,利用支持向量机对现勘图像进行分类从而得到高效的图像特征。对比实验结果表明,该算法的分类准确率较高。

To deal with the problem of the low accuracy of the crime scene investigation image classification(CSIC)due to imprecise description of the image content using the low level features,an improved LLC(Local-Constrained Linear Coding)algorithm using deep learning features to obtain efficient image feature representation is proposed in this paper.In this algorithm,the dense local features of images based on the trained convolutional neural networks is extracted using sliding window method.Then the approximated LLC method is used to fast encoding and max pooling the dense CNN features.The multiscale spatial pyramid matching is further used to generate sparse coding features containing spatial location information.Finally,SVM(Support Vector Machine)classifier is exploited for object classification.Experimental results on several image dataset show that the proposed method can achieve better performance than the state-of-the-art image classification methods.