研究目的
Investigating the therapeutic effects of a specific herbal medicine on a particular disease.
研究成果
The proposed technique effectively segments cracks from EL images of mono- and poly-crystalline solar cells with just one forward pass through the network, given only a small number of training samples and only image-level annotations. The study of normalized Lp normalization with different values of p provides a basis for the design of new weakly supervised segmentation procedures.
研究不足
The technical and application constraints include the limited amount of data and the challenge of segmenting cracks on electroluminescence images of mono- or polycrystalline solar modules. Potential areas for optimization include the choice of the best norm for the given problem and the inversion of intensities for lower norms.
1：Experimental Design and Method Selection:
The methodology involves a weakly supervised learning strategy using a modified ResNet-50 for segmentation from network activation maps, with defect classification as a surrogate task. Normalized Lp normalization is applied to aggregate activation maps into single scores for classification.
2：Sample Selection and Data Sources:
The dataset consists of 2426 electroluminescence 8-bit grayscale images of solar cells with a resolution of 300×300 pixels per image, including expert annotations of the probability that a cell contains any kind of defect.
3：List of Experimental Equipment and Materials:
A modified ResNet-50 architecture is used, initialized with pretrained weights trained on ImageNet.
4：Experimental Procedures and Operational Workflow:
The network is trained using categorical cross entropy as loss function, with modifications to preserve spatial information and increase resolution of segmentation masks.
5：Data Analysis Methods:
The segmentation is performed by pixel-wise comparing the heatmap value with the half of the maximum value of the heatmap segment.
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
