摘要： 为提高低配置计算环境中的视觉目标实时在线分类特征提取的时效性和分类准确率,提出一种新的目标分类特征深度学习模型。根据高时效性要求,选用分类器模型离线深度学习的策略,以节约在线训练时间。针对网络深度受限和高识别率要求,提取图像的局部方向梯度直方图(HOG)特征,构建稀疏自编码器栈对HOG特征进行深层次编码,设计Softmax多分类器对所抽取的特征进行分类。在深度神经网络模型学习过程中,引入最小化各层结构风险和微调全网参数的二阶段最优化策略。利用场景图像库Caltech101和手写数字库MNIST的训练样本与测试样本进行对比实验,结果表明,该模型在局部特征提取方面的时效优于单层卷积神经网络(CNN)模型,分类准确率高于CNN、栈式自编码器等对比模型。

关键词: 计算机视觉, 目标分类, 方向梯度直方图特征, 栈式自编码器, 深度学习

Abstract: To improve the feature extraction timeliness and classification validity of real-time classification of visual object in low computing profile,a Histogram of Oriented Gradients(HOG)-based feature deep learning model for object classification is proposed.For the requirements of high timeliness,offline deep learning strategy is applied to the classifier model to save its online training time.In view of the requirements of network depth limitation and high recognition rate,the local feature of HOG feature of an image is extracted to be used as the input of the sparse autoencoder stack so as to output the high level feature code of the sample image.The Softmax multiple classifier is designed to classify the extracted features.During the learning process of the deep neural network model,the two-stage optimization strategy is introduced,which minimizes the structural risk of every layer and fine-tune the parameters of the whole model.Using some samples of the scene image database Caltech101 and that of the handwritten digits database MNIST as the training set and the others as the test set to perform the comparative experiment,results show that the time performance of the proposed model is better than that of one-layer only Convolutional Neural Network(CNN),and the classification accuracy of the trained model is higher than that of CNN,Stacked Autoencoder(SAE) comparative models.

Key words: computer vision, object classification, Histogram of Oriented Gradients(HOG) feature, Stacked Autoencoder(SAE), deep learning

中图分类号: 

• TP391