Explore chapters and articles related to this topic
Intelligent Systems
Published in Puneet Kumar, Vinod Kumar Jain, Dharminder Kumar, Artificial Intelligence and Global Society, 2021
Satyajee Srivastava, Abhishek Singh, Deepak Dudeja
Triplet loss is a kind of loss function used in neural networks. Gradient descent can be applied on a triplet loss function. A triplet loss function is simply considered as a loss function using three images: first is an anchor image, second is a positive image P (same Figure 16.11 person as the anchor), and third is a negative image (another person from the anchor). It has a baseline as an input (an anchor) (as shown in Figure 16.11). It is compared with a positive input and a negative input [28]. The triplet loss minimizes the distance between an anchor (baseline) and a positive — having the same identity — and maximizes the distance between anchor (baseline) and a negative — having different identities.
Deep Face Recognition
Published in Hassan Ugail, Deep Learning in Visual Computing, 2022
Common FaceNet models use two types of architectures. They are the Zeiler and Fergus architecture and GoogLeNet style Inception model. The essential idea in the training of the FaceNet is the “triplet loss” to capture the similarities and differences between classes of faces in a 128-dimensional embedding. Hence, given an embedding E(x) from an image to a feature set Rn, FaceNet looks into the squared L2 distance between face images such that this value is small for images of the same identity and is large for the different identities.
Real-Time Identity Censorship of Videos to Enable Live Telecast Using NVIDIA Jetson Nano
Published in Mohamed Lahby, Utku Kose, Akash Kumar Bhoi, Explainable Artificial Intelligence for Smart Cities, 2021
Shree Charran R., Rahul Kumar Dubey
The learning objective of MTCNN is a multi-task loss, with one loss as binomial cross-entropy loss is the probability that the box has a face, the second one is Euclidean distance loss for bounding box regression, and third one is Euclidean loss for facial landmark regression. The 3 losses are weighted and summed up in a cumulative multi-task formula. The output bounding box with the face from MTCNN is fed to FaceNet for face recognition in the bounding box. FaceNet is a face recognition model by Google. The FaceNet system can be used to extract high-quality features from faces, called face Embedding’s that can then be used to train a face identification system. Finally, an SVM classifier is used to identify the face in the last stage. The face Embedding’s are multidimensional numerical vector representations of a face which represent the unique identity of the face. Face net provides a 128 dimension embedding for a face. The triplet loss involves comparing face Embedding’s for three images, one being an anchor (reference) image, second, a positive image (matching the anchor), and third, a negative image (not matching the anchor). Embedding’s are learnt by a Deep CNN network such that the positive embedding is closer to anchor embedding compared to negative embedding distance to the anchor F(A)−F(P)+margin<F(A)−F(N) where
Semi-supervised pipeline anomaly detection algorithm based on memory items and metric learning
Published in Nondestructive Testing and Evaluation, 2023
Bingchuan Yan, Jianfeng Zheng, Rui Li, Kuan Fu, Pengchao Chen, Guangming Jia, Yunhan Shi, Junshuang Lv, Bin Gao
The triplet loss function is determined by four parameters: anchor (), positive (), negative ()and margin (). The goal of triplet loss is to make the features of the same label (anchor and positive) as close as possible and the features of different labels (anchor and negative) as far as possible in feature space. At the same time, in order to prevent the samples’ feature from converging into a very small space, it is required that the negative case should be at least margin away from the positive case. We consider there are two different type image, normal sample and abnormal sample, in the training dataset. So, we chose different vectors as negative parameter in the training stage.
LifeGuard: An Improvement of Actor-Critic Model with Collision Predictor in Autonomous UAV Navigation
Published in Applied Artificial Intelligence, 2022
Manit Chansuparp, Kulsawasd Jitkajornwanich
There are various improvements on autoencoder such as denoising autoencoder (Vincent et al. 2008), which add noise on the input to make the network more robust against discrepancy, sparse autoencoder (Ng 2011), which regularize autoencoder by sparsity constraint, and so on. In addition, there is a combination between autoencoder and Siamese network model with triplet loss (Schneider et al. 2019) to exploit the good representation in latent space and distance-based classification ability. The triplet loss function helps the features of data in the same class get closer than those from other classes. The function can be defined as follows: