电光与控制, 2019, 26 (9): 90, 网络出版: 2021-01-31   

基于CNN的不平衡SAR图像舰船目标识别

CNN Based Ship Target Recognition of Imbalanced SAR Image
作者单位
海军航空大学, 山东 烟台 264001
摘要
针对SAR图像舰船目标识别中存在的数据不平衡问题提出了一种基于批内平衡采样和模型微调的两阶段迁移学习方法。首先使用批内平衡采样方法得到类间数量平衡的训练集, 然后使用该数据集预训练模型, 最后通过迁移学习和模型微调继续训练不平衡数据并完成测试。针对一般三通道CNN模型在处理单通道SAR图像时会出现参数冗余的问题,设计了一种用于SAR图像识别的轻量化CNN模型。通过单通道卷积核、深度可分离卷积和用全局平均池化代替全连接层3种策略有效降低了模型的参数量。在公开数据集OpenSARShip上的实验结果表明: 所提方法有效提升了少数类的识别精度, 缓解了数据不平衡问题对识别结果的影响; 所提轻量化CNN模型在保证识别精度基本不变的前提下, 使传统三通道CNN模型的模型大小和单次迭代时间分别降低约58.86%和63.62%。
Abstract
Aiming at data imbalance in SAR image recognitiona two-stage transfer learning method based on intra-batch balanced sampling and model fine tuning is proposed. First, the training set of balance data between classes is obtained by using the intra-batch balanced sampling method. Then the training data set is used for pre-training of the model. Finally, the training of imbalanced data and the test is completed by transfer learning and model fine tuning. Aiming at the problem of redundant parameters in the processing of single channel SAR images by using general tri-channel CNN model, a lightweight CNN model is designed for SAR image recognition. Through the three strategies of single-channel convolution kernel, deep separable convolution and using global average pooling instead of full connected layer, the parameters of the model are reduced greatly. The results of experiments conducted on the open data set OpenSARShip show that: 1) The proposed method effectively improves the recognition accuracy of minority classes and reduces the effect of data imbalance on recognition results;and 2) The proposed lightweight CNN model can reduce the size and the single-iteration time of the traditional tri-channel CNN model by about 58.86% and 63.62% respectively, under the premise that the recognition accuracy is basically unchanged.

邵嘉琪, 曲长文, 李健伟, 彭书娟. 基于CNN的不平衡SAR图像舰船目标识别[J]. 电光与控制, 2019, 26(9): 90. SHAO Jiaqi, QU Changwen, LIJianwei, PENG Shujuan. CNN Based Ship Target Recognition of Imbalanced SAR Image[J]. Electronics Optics & Control, 2019, 26(9): 90.

本文已被 6 篇论文引用
被引统计数据来源于中国光学期刊网
引用该论文: TXT   |   EndNote

相关论文

加载中...

关于本站 Cookie 的使用提示

中国光学期刊网使用基于 cookie 的技术来更好地为您提供各项服务,点击此处了解我们的隐私策略。 如您需继续使用本网站,请您授权我们使用本地 cookie 来保存部分信息。
全站搜索
您最值得信赖的光电行业旗舰网络服务平台!