Research Publications/Outputs

Permanent URI for this community

http://hdl.handle.net/10204/11481

Browse

Browsing Research Publications/Outputs by browse.metadata.impactarea "Artificial Intel Augment Real"

Now showing 1 - 2 of 2
  • Thumbnail Image
    Item
    Age invariant face recognition methods: A review
    (2021-12) Baruni, Kedimotse P; Mokoena, Nthabiseng ME; Veeraragoo, Mahalingam; Holder, Ross P
    Face recognition is one of the biometric technologies that is mostly used in surveillance and law enforcement for identification and verification. However, face recognition remains a challenge in verifying and identifying individuals due to significant facial appearance discrepancies caused by age progression. Especially in applications that verify individuals from their passports, driving licenses and finding missing children after decades. The most critical step in Age- Invariant Face Recognition (AIFR) is extracting rich discriminative age-invariant features for each individual in face recognition applications. The variation of facial appearance across aging can be solved using three methods, namely, generative (aging simulation), discriminative (feature-based) and deep neural networks methods. This work reviews and compares the state-of-art AIFR methods to address the work that has been done to minimize the effect of aging in face recognition application during the pre-processing and feature extraction stages to extract rich discriminative age-invariant features from facial images of individuals (subjects) captured at different ages, shortfalls and advantages of these methods. The novelty of this work lies in analyzing the state-of-art work that has been done during the pre-processing and/or feature extraction stages to minimize the difference between the query and enrolled face images captured over age progression.
  • Thumbnail Image
    Item
    A benchmark dataset for defect detection and classification in electroluminescence images of PV modules using semantic segmentation
    (2023-12) Pratt, Lawrence E; Matheus, Jana; Klein, R
    Electroluminescence (EL) images enable defect detection in solar photovoltaic (PV) modules that are otherwise invisible to the naked eye, much the same way an x-ray enables a doctor to detect cracks and fractures in bones. This paper presents a benchmark dataset and results for automatic detection and classification using deep learning models trained on 24 defects and features in EL images of crystalline silicon solar cells. The dataset consists of 593 cell images with ground truth masks corresponding to the pixel-level labels for each feature and defect. Four deep learning models (U-Net_12, U-Net_25, PSPNet, and DeepLabv3+) were trained using equal class weights, inverse class weights, and custom class weights for a total of twelve sets of predictions for each of 50 test images. The model performance was quantified based on the median intersection over union (mIoU) and median recall (mRcl) for a subset of the most common defects (cracks, inactive areas, and gridline defects) and features (ribbon interconnects and cell spacing) in the dataset. The mIoU measured higher for the two features compared to the three defects across all models which correlates with the size of the large features compared to the small defects that each class occupies in the images. The DeepLabv3+ with custom class weights scores the highest in terms of mIoU for the selected defects in this dataset. While the mIoU for cracks is low (25%) even for the DeepLabv3+, the recall is high (86%), and the resulting prediction masks reliably locate the defects in complex images with both large and small objects. Therefore, the model proves useful in the context of detecting cracks and other defects in EL images. The unique contributions from this work include the benchmark dataset with corresponding ground truth masks for multi-class semantic segmentation in EL images of solar PV cells and the performance metrics from four semantic segmentation models trained using three sets of class weights.