基于深度卷积神经网络的羽绒图像识别-《郑州大学学报(工学版)》

文章信息/Info

Author(s):: Yang Wenzhu; Liu Qing; Wang Sile; Cui Zhenchao; Zhang Ningyu; School of Cyberspace Security and Computer, Hebei University, Baoding, Hebei, 071002

Keywords:: deep convolutional neural networks; weights initialization; sparse autoencoder; visual saliency; image recognition

摘要:: 由于图像中羽绒形态及其多样性，传统的图像识别方法难以正确识别羽绒分拣图像中的羽绒类型，其识别精度也难以达到实际生产的要求．为解决上述问题，构造了一种用于羽绒类型识别的深度卷积神经网络，并对其权值初始化方法进行了改进．首先利用视觉显著性模型提取羽绒图像的显著部分，然后将图像的显著部分输入到稀疏自动编码器中进行训练，得到一组符合数据集统计特性的卷积核集合．最后采用Inception及其变种模块实现深度卷积神经网络的构造，通过增加网络深度来提高网络的识别精度．试验结果表明，用所构造的深度卷积神经网络对羽绒图像识别的精度较传统卷积神经网络提高了2.7%，且改进的权值初始化方法使网络的收敛速度提高了25.5%．

Abstract:: Bcause of the scale and the various shapes of down in the image, it was difficult for traditional image recognition method to correctly recognize the type of down image and got the required recognition accuracy, even for the Traditional Convolutional Neural Networks (TCNN). To solve the above problems, a Deep Convolutional Neural Networks (DCNN) for down image recognition was constructed, and a new weight initialization method was proposed. Firstly, these salient regions of images were cut from the images using the visual saliency model.Then, these salient regions were used to train a sparse autoencoder and get a collection of convolutional filters, which accord with the statistical characteristics of dataset. At last, a DCNN with Inception module and its variants was constructed. To enhance the recognition accuracy, the depth of the network was deepened. The experiment results indicated that the constructed DCNN increased the recognition acuracy by 2.7% compared to TCNN, when recognizing the down in the images. The convergence rate of the proposed CNN with the new weight initialization method was improved by 25.5% compared to TCNN.