研究目的
To address the problem of over-fitting in training convolutional neural networks (CNNs) for ship classification in synthetic aperture radar (SAR) images with small datasets by proposing a new data augmentation method combined with transfer learning.
研究成果
The proposed method of data augmentation combined with transfer learning effectively addresses over-fitting in small SAR datasets, achieving high classification accuracies (up to 98.96% with Densenet-121) and outperforming existing methods. It demonstrates good generalization ability and is applicable to other SAR fields with limited data, though future work could explore additional preprocessing techniques like GAN or SRCNN.
研究不足
The study uses a small dataset of 250 images, which may limit generalizability. The proposed data augmentation method may introduce noise, and the effectiveness could vary with different SAR image characteristics or larger datasets. Computational resources and specific hardware are not detailed, potentially affecting reproducibility.
1：Experimental Design and Method Selection:
The study uses CNNs for ship classification, specifically employing traditional CNN models and Resnet models with transfer learning. A new data augmentation method is proposed to increase dataset size while retaining important information, involving rotation with pixel padding to avoid black areas.
2：Sample Selection and Data Sources:
The dataset consists of 250 SAR ship images from TerraSAR-X stripmap mode, with resolutions of 2×
3：5 m, including 150 Bulk Carrier, 50 Container Ship, and 50 Oil Tanker images. Images are manually annotated using AIS data. List of Experimental Equipment and Materials:
SAR images from TerraSAR-X, computational resources for deep learning (implied but not specified).
4：Experimental Procedures and Operational Workflow:
Data is split into training (70%) and validation (30%) sets. Three datasets (D1, D2, D3) are created using different augmentation methods: D1 with traditional methods (flipping, brightening, etc.), D2 with random crop, and D3 with the proposed rotation method. Models are trained using SGD with momentum, a minibatch size of 64, weight decay of
5：00000001, initial learning rate of 001 reduced by 1 after 100 epochs, and momentum of Data Analysis Methods:
Performance is evaluated using accuracy and f1-score, with comparisons to other methods and models.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
