5 research outputs found
Improving Noisy Face Embeddings Using Time Series Transformers and Adaptive Triplet Loss
No full textFace recognition accuracy is now at the point that it is comparable to human performance, largely driven by improvements in face embedding models. However, face embedding models rely on high-quality images, making them vulnerable to variations in quality, illumination, pose, and occlusion in real-world scenarios. To address this, we propose a Transformer-based aggregation method for face verification that leverages multiple embeddings to improve the robustness under adverse conditions. The model refines frame-level embeddings computed by a high-performance single-frame feature extractor using a Transformer encoder and is trained with an adaptive triplet loss that applies separate adaptive margins for positive and negative pairs. We train the embedding aggregation model with a two-phase strategy: pretraining with average pooling alignment and fine-tuning with adaptive triplet loss. We train the embedding aggregation model on the YouTube Faces dataset and evaluate it on unseen data from IMFDB, CASIA-WebFace, CCVID, IJB-B, and IJB-C, where it outperforms the baseline on most datasets, especially IMFDB, which has significant variations in pose, lighting, and scene context. Our approach is an effective way to build upon a single-frame face embedding model, leveraging multiple images for accurate face verification under adverse conditions
Convolutional Neural Networks for Herb Identification: Plain Background and Natural Environment
Get PDFConvolutional neural networks have achieved success in resolving object identification problems. This study contributes a suitable new approach to herb identification for educational and research purposes based on a small dataset and small-sized images. Two self-collected Thai herb datasets with either plain or natural environment backgrounds were used for experimentation to realize this objective. The plain background dataset includes 4,400 images of 11 leaf types, and the natural dataset contains 1,620 images of nine leaf types. The images were divided into a training set containing 75% of the images and a separate test set with the remaining 25%. The experiments included five-fold cross-validation applied to the training set; the InceptionV3, MobileNetV2, ResNet50V2, VGG16, and Xception convolutional neural network models RMSprop and Adam optimizers. Further, dropout rates of 0.3, 0.5, and 0.7 were considered along with five and ten epochs. Transfer learning was applied using pre-trained weights. The model with the best outcome, based on the average accuracy of the cross-validation results on both datasets (the plain background dataset was 94.55%, and the natural dataset was 90.37%), was the VGG16 with the RMSprop optimizer, which exhibited a dropout rate of 0.5 over ten epochs. The model achieved 96.64% and 92.00% accuracy on the plain background training and test sets, and 99.59% and 91.36% on the natural environment training and test sets, respectively. The results show that the method has a high potential for objective tasks and application in identifying herbs based on visual leaf information
