光学 精密工程, 2020, 28 (6): 1404, 网络出版: 2020-06-04
循环神经网络多标签航空图像分类
Recurrent neural network multi-label aerial images classification
航空图像分类 多标签 注意力机制 多尺度 卷积神经网络 长短期记忆网络 satellite images classification muilti-label attention mechanisms multi-scale convolutional neural network Long Short-Term Memory(LSTM) network
摘要
由于航空图像背景复杂,包含的物体类别多样,航空图像分类任务仍然面临困难。针对传统航空图像多标签分类算法准确率低、泛化性差的问题,本文提出了一种基于循环神经网络多标签航空图像分类方法。首先,采用超像素分割获取图像的低层特征,通过注意力机制生成注意力特征图;接着,采用交叉验证的方式获取最佳的图像尺度,将多尺度注意力特征图嵌入卷积神经网络中对 图像进行特征提取;最后,采用改进的双向长短期记忆网络挖掘标签之间的相关性,改进的双向长短期记忆网络增加了输入门到输出门之间的连接,使输入状态可以更好地控制每一内存单元输出的 信息,并且将遗忘门和输入门合并成单一的更新门,使得改进的双向长短期记忆网络可以学到更长时期的历史信息。结果显示,在图像变换尺度为1,1.3,2时,模型在UCM多标签数据集上的精确率 和召回率分别达到了85.33%和87.05%,F1值达到了0.862。本文方法相比于原始VGGNet16模型,精确率提高了7.25%,召回率提高了8.94%。实验表明,该方法可以有效提高航空图像多标签分类任务 的准确率。
Abstract
Due to the complexity of the background in aerial images and the diversity of object categories, aerial image classification is a challenging task. In order to address the problems of low accuracy and poor generalization in traditional multi-label aerial image classification methods, a method based on recurrent neural networks was proposed.In this method, the super-pixel segmentation algorithm was first used to obtain the low-level features of the image from which an attention map was generated. Subsequently, the best image scale was obtained by cross-validation,and multi-scale attention feature graphs were embedded into aconvolutional neural network in order to extract the features of the image.Finally, tomine the correlation between labels,an improved bidirectional Long Short-Term Memory (LSTM)network was proposed, which increases the connection from the input gate to the output gate, so that the input state can efficiently control the output information of each memory unit. The forget gate and the input gate were combined into a single update gate so that the improved bidirectional LSTM network can learn long-term historical information. The results obtained by applying the proposed method to the UCM multi-label dataset indicate that for scale values of 1,1.3, and 2, the accuracy and recall rates of the model are 85.33% and 87.05% respectively,while the F1 score reached 0.862. The accuracyand recall rates are found to be higher than those of theVGGNet16 model by 7.25% and 8.94% respectively.The experimental results thus indicate that the proposed method can effectively increase the accuracy of multi-label aerial image classification.
陈科峻, 张叶. 循环神经网络多标签航空图像分类[J]. 光学 精密工程, 2020, 28(6): 1404. CHEN Ke-jun, ZHANG Ye. Recurrent neural network multi-label aerial images classification[J]. Optics and Precision Engineering, 2020, 28(6): 1404.