Development history and commonly used data sets of face recognition

Early Stages of Face Recognition - Based on Machine Learning

Early methods mainly focused on working with computer vision experts to extract hand-crafted features, And use traditional machine learning algorithms to train effective classifiers for detection. However, the limitations of these methods are that experts are required to produce effective features, and each component needs to be optimized individually, resulting in the entire detection pipeline being under-optimized. To solve this problem, more complex features such as HOG, SIFT, SURF and ACF have been proposed. To enhance the robustness of detection, combinations of multiple detectors trained for different views or poses have also been developed. However, these models require long training and testing times and have limited improvement in detection performance.

More advanced technology for face recognition - based on deep learning

In recent years, research on facial recognition has made significant progress, especially It is an application of deep convolutional neural network (CNN). Deep learning methods have achieved remarkable success in computer vision tasks and have many advantages over traditional methods. Deep learning methods avoid handcrafted design pipelines, which makes models more flexible and adaptable to different data sets. Furthermore, deep learning methods have performed well in many benchmark evaluations, such as the ImageNet Large Scale Visual Recognition Challenge (ILSVRC). These advances have enabled facial recognition to be widely used in various fields, from security surveillance to face unlocking.

Recently, researchers have made exciting progress in the field of general object detection using Faster R-CNN, an advanced object detector. By combining the joint training of CNN cascade, region proposal network (RPN) and Faster R-CNN, the researchers achieved end-to-end optimization and achieved encouraging results. In terms of face detection, the Faster R-CNN algorithm is combined with hard negative mining and ResNet, which greatly improves its performance on face detection benchmarks such as FDDB. This combined approach makes the face detection algorithm more accurate and reliable. In short, Faster R-CNN and its related joint training and combination algorithms have brought significant progress to the fields of object detection and face detection, and opened up a new direction for the development of deep learning technology.

Commonly used datasets for face recognition

AFW dataset: The AFW dataset is built using Flickr images. It includes 205 images and 473 labeled faces. For each face, image annotations include a rectangular bounding box, 6 landmarks, and pose angles.

PASCAL FACE dataset: This dataset is used for facial recognition and face recognition; it is a subset of PASCAL VOC and contains 851 images with large facial appearance and pose variations. The image contains 1335 labeled faces.

MIT CBCL face database: The MIT-CBCL face recognition database contains a training set (2429 faces, 4548 non-faces) and an estimation set (472 faces, 23573 a non-human face).

FDDB Dataset: This dataset contains 5171 faces with annotations such as occlusions, difficult poses and low image resolution in 2845 images. These images are used for training on large appearance variations, severe occlusions, and severe blur degradation, which are common when detecting faces in unconstrained real-life scenarios.

CMU PIE database: The CMU Multi-PIE Face database contains 41,368 images of 68 people, each person in 13 different poses, 43 different lighting conditions and 4 different expression.

SCface dataset: SCface is a face static image database. The images were captured using five video surveillance cameras of varying quality in an uncontrolled indoor environment. This dataset contains 4160 static images (visible and infrared spectra) of 130 subjects.

WIDER FACE dataset: The face detection benchmark dataset includes 32,203 images and 393,703 labeled faces with high variability in scale, pose, and occlusion, This makes face detection extremely challenging. Furthermore, the WIDER FACE dataset is organized according to 61 event categories.

The above is the detailed content of Development history and commonly used data sets of face recognition. For more information, please follow other related articles on the PHP Chinese website!