研究目的
To address the problem of false alarms in ship detection from SAR images caused by azimuth ambiguity and buildings exhibiting similar scattering mechanisms, by applying self-designed deep convolutional neural networks to discriminate between true ships and false alarms.
研究成果
The self-designed deep convolutional neural network effectively discriminates ships from false alarms in SAR images, achieving over 95% classification accuracy on both datasets. This demonstrates the capability of deep learning to automatically learn discriminative features for SAR-based ship discrimination, with significant improvements over existing methods, particularly in true positive classification. Future work will focus on feature analysis and additional evaluations.
研究不足
The model requires feature analysis of ships and further evaluation on the second dataset, which are planned for future work. Potential optimizations include deeper analysis of misclassifications and extension to more diverse datasets.
1：Experimental Design and Method Selection:
The study uses a self-designed deep convolutional neural network (CNN) inspired by VGG16, with modifications for SAR image processing, including 'same' padding and specific layer configurations. The methodology involves training and validation processes to classify ship chips into categories such as true positive, false positive, and true negative.
2：Sample Selection and Data Sources:
Two datasets are used: one reconstructed from IEEEDataPort SARSHIPDATA (with 22 dual Sentinel-1 and 2 single TerraSAR-X images, providing ship chips of size 48x48 pixels in three categories) and another constructed from 10 scenes of Sentinel-1 SAR images (with two classes: ship and not-ship chips). Data is normalized to [0,255] and stored in JPG format.
3：List of Experimental Equipment and Materials:
The implementation uses Tensorflow on Ubuntu 14.04 with one GTX 1070 GPU (8GB memory).
4：04 with one GTX 1070 GPU (8GB memory).
Experimental Procedures and Operational Workflow:
4. Experimental Procedures and Operational Workflow: Ship chips are acquired from SAR images, split into training (80%) and validation (20%) datasets. The CNN is trained using the Adam optimizer with hyperparameters: learning rate 0.001, dropout probability 0.5, batch size 256. Training involves 150 epochs, and the model is evaluated on validation data.
5：001, dropout probability 5, batch size Training involves 150 epochs, and the model is evaluated on validation data.
Data Analysis Methods:
5. Data Analysis Methods: Classification accuracy and loss are monitored during training and testing. Results are compared with other methods (Cubic SVM, Highway-50 network, Adaboost) using accuracy metrics.
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
