研究目的
Investigating the efficiency and accuracy of a novel end-to-end deep learning-based architecture for defects segmentation in photovoltaic electroluminescence images.
研究成果
The proposed method segments defects accurately with less false detection than state-of-the-art methods, achieving a mean IOU of 0.6477 and pixel accuracy of 0.9738. The global attention network enhances defect features, and dilation extracts large semantic information. Future work will address the sensitivity drawback.
研究不足
The proposed network has less sensitivity than state-of-the-art methods, possibly due to the lack of pre-processing or post-processing techniques.
1：Experimental Design and Method Selection:
The study proposes a novel end-to-end deep learning-based architecture for defects segmentation in photovoltaic electroluminescence images, incorporating a novel global attention mechanism and modified U-net with dilated convolution and skip connections.
2：Sample Selection and Data Sources:
The dataset consists of 400 defected 512x512 EL images from a real industrial environment, with 350 images used for training and 50 for testing.
3：List of Experimental Equipment and Materials:
The implementation uses Keras 2.2.4 framework, trained on GTX 1080 GPU, with Intel? Core? i7-6700K CPU @ 4.00 GHZ, 32GB RAM, and 2TB hard disk.
4：4 framework, trained on GTX 1080 GPU, with Intel? Core? i7-6700K CPU @ 00 GHZ, 32GB RAM, and 2TB hard disk.
Experimental Procedures and Operational Workflow:
4. Experimental Procedures and Operational Workflow: The model is trained for 100 epochs with a batch size of 1, using RmsProp optimizer with a learning rate of 0.0001 and rho=0.9. Data augmentation techniques include rotation, width shift, height shift, shear range, horizontal flips, vertical flips, and adaptive histogram equalization.
5：0001 and rho=Data augmentation techniques include rotation, width shift, height shift, shear range, horizontal flips, vertical flips, and adaptive histogram equalization.
Data Analysis Methods:
5. Data Analysis Methods: Performance is evaluated using Intersection over union (IOU), sensitivity, and specificity 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
