计算机科学
残差神经网络
人工智能
上下文图像分类
模式识别(心理学)
图像(数学)
计算机视觉
深度学习
作者
Zhengdi Sima,Jingyu Tao,Zhaochen Liu
摘要
Image classification is one of the important basic tasks in computer vision. As the basic module of various complex tasks, ResNet of various types plays an important role as a back- bone for various tasks. This paper aims to use a series of effective methods to improve the performance of traditional ResNet and, to some extent, improve its generalization ability. We improve the methods of ResNet in terms of refined image conversion, optimizer improvement, adaptive learning rate scheduling, difference classification weighting, and experimentally using validation set integration, as well as in the structure of ResNet itself by incorporating residual structures in the network by using ResNet. With these bar improvements, our new baseline performance has made a significant difference, and we also have excellent performance in generalization migration to multiple tasks.
科研通智能强力驱动
Strongly Powered by AbleSci AI