摘要
图像标注作为有监督学习的一个典型应用,一直深受研究者的关注。图像标注模型中度量图像样本损失函数的合适选取,对提升图像标注模型的预测准确率,具有重要的指导作用。从分析损失函数对模型预测性能影响的角度出发,首先对基于神经网络的单标签图像标注方法,在MNIST数据集下,通过更换神经网络模型的损失函数,对比研究了有监督学习中常用损失函数度量样本的性能差异,然后给出了一种新的损失函数,最后实验验证了该损失函数的有效性。为有监督学习算法中损失函数的有效构造,提高图像标注性能提供了一种思路。
As a typical application of supervised learning,image annotation has been deeply concerned by researchers.The proper selection of image sample loss function in image annotation model has an important guiding role in improving the prediction accuracy of the image annotation model.From the perspective of analyzing the impact of loss function on the model prediction performance,the loss function of the neural network model is replaced in single-label image annotation method under the MNIST data set.Then the difference performance of common loss function on samples has been comparatively studied in supervised learning,and a new loss function is presented.Finally,its effectiveness is proved and a way is provided for constructing the loss function and improving the image annotation performance.
作者
邓建国
张素兰
DENG Jian-guo;ZHANG Su-lan(School of Computer Science and Technology,Taiyuan University of Science and Technology,Taiyuan 030024,China)
出处
《太原科技大学学报》
2020年第6期433-439,448,共8页
Journal of Taiyuan University of Science and Technology
基金
国家自然科学基金(61373099)。
关键词
有监督学习
损失函数
图像标注
supervised learning
loss function
image annotation
