激光与光电子学进展, 2019, 56 (23): 231502, 网络出版: 2019-11-27
一种多任务的卷积神经网络目标分类算法 下载: 1076次
Object Classification Based on Multitask Convolutional Neural Network
机器视觉 多任务卷积神经网络 深度学习 细粒度图像 machine vision multi-task convolutional neural network deep learning fine-grained image
摘要
提出一种基于细粒度图像和多属性融合的多任务卷积神经网络(MTCNN)。该网络主要包含几个关键环节,首先在网络中增加标签输入层,复制并分离输入的多个标签,通过全连接层与多个任务相匹配,增加与标签数量相应的Softmax损失函数,来对多个任务进行反向传播;然后,使用显著性检测与角点检测相结合的方法,提取出原始图像中的细粒度图像用于MTCNN的数据输入,使神经网络提取到的目标特征具有独特性和区分性;最后,使用非线性激活函数PReLu,进一步提高网络的分类精度。通过在Car Dataset中进行多任务并行训练,测试精度较传统单个任务的分类精度提升10%,实验结果表明,MTCNN有较高的泛化能力,对于图像分类的精度有明显的提升。
Abstract
This paper proposes a multitask convolutional neural network (MTCNN) based on fine-grained images and multi-attribute fusion. The network mainly includes the following key links. First, the label input layer is added to the network; the input multiple labels are copied and separated, and then matched to multiple tasks with a fully connected layer. The Softmax loss function corresponding to the number of labels is added to backpropagate multiple tasks. Then, a fine-grained image in the original image is extracted by the combination of saliency detection and corner detection, and used as the input of MTCNN. The target features extracted by the neural network are more unique and distinguishable. Finally, the MTCNN uses the nonlinear activation function PReLu to further improve the classification accuracy of the network. This paper uses the MTCNN to perform multi-task parallel training in the Car Dataset and achieves a 10% improvement in the classification accuracy over the traditional single task. The results show that the MTCNN has high generalization performance and the accuracy of image classification is obviously improved.
张苗辉, 张博, 高诚诚. 一种多任务的卷积神经网络目标分类算法[J]. 激光与光电子学进展, 2019, 56(23): 231502. Miaohui Zhang, Bo Zhang, Chengcheng Gao. Object Classification Based on Multitask Convolutional Neural Network[J]. Laser & Optoelectronics Progress, 2019, 56(23): 231502.