基于特征融合的细粒度鸟类图像分类研究
Research on fine-grained bird image classification based on feature fusion
李昊霖 1俞成海 1卢智龙 1陈涵颖2
作者信息
- 1. 浙江理工大学计算机科学与技术学院,浙江 杭州 310018
- 2. 华北电力大学
- 折叠
摘要
特征金字塔(FPN)因能将低尺度的特征与更高尺度的特征融合、呈现每个层次丰富的语义信息,而被广泛应用于小尺度目标定位识别中,但其目前无法连接跨尺度特征信息,且分类准确率不高.本文提出特征融合金字塔模块(FFPN),通过在ResNet50主干网络中引入FFPN模块,有效地提高了细粒度鸟类图像分类的性能.模型在CUB-200-2011数据集上达到了83.379%的分类准确度,在Bird-400数据集中达到了91.201%的准确度,实现了较好的分类效果.
Abstract
Feature pyramid network(FPN)is widely used for small object detection and localization,owing to its ability to fuse features from different scales to provide rich semantic information for each feature level.However,the current FPN still cannot build connections between features across scales,and has suboptimal classification accuracy.To address this,the feature fusion pyramid network(FFPN)is proposed,which effectively improves the performance of fine-grained bird image classification by incorporating FFPN modules into the ResNet50 backbone.The model achieves 83.379%classification accuracy on CUB-200-2011 dataset and 91.201%on Bird-400 dataset,realizing good classification results.
关键词
特征融合/多尺度特征/细粒度图像分类/鸟类图像识别Key words
feature fusion/multi-scale features/fine-grained image classification/bird image recognition引用本文复制引用
出版年
2023
NETL
NSTL
维普
万方数据