"深度人脸识别综述：最新进展与关键元素"

需积分: 11 40 浏览量更新于2024-03-22 收藏 9.28MB PDF 举报

In their paper "The Elements of End-to-end Deep Face Recognition: A Survey of Recent Advances," Hang Du, Hailin Shi, Dan Zeng, and Tao Mei provide a comprehensive overview of the advancements in deep face recognition. Face recognition has always been a key focus in the field of computer vision, and with the advent of deep convolutional neural networks and the availability of massive datasets, significant strides have been made in this area. The authors delve into the key elements of end-to-end deep face recognition, highlighting the various techniques and algorithms that have contributed to the advancements in this field. They discuss the importance of features extraction, data augmentation, and model training in achieving state-of-the-art results in face recognition. Additionally, they explore the challenges and potential solutions for face recognition in real-world applications. The paper not only provides a thorough examination of the current state of deep face recognition but also offers insights into future research directions. The authors emphasize the importance of continual research and development in this area to address the evolving needs and challenges of face recognition technology. Overall, "The Elements of End-to-end Deep Face Recognition: A Survey of Recent Advances" serves as a valuable resource for researchers and practitioners in the field of computer vision. It sheds light on the progress that has been made in deep face recognition and points towards exciting possibilities for future advancements in this rapidly growing field.

Fig. 5. The illustration of single-stage and multi-stage face detectors. The single-stage detector directly

accomplishes the face detection from the entire feature maps, whereas the multi-stage detector adopts a

proposal stage to generate candidates and one or more stages to refine these candidates.

Apart from the modeling, how the train the multi-stage detector is another interesting topic.

The multi-stage detectors are commonly trained stage by stage, since each stage is supervised

by its own objective. This may lead to inferior optimization. To handle this issue, a joint training

strategy [

178

] was designed for both Cascaded CNN [

123

] and Faster R-CNN to achieve end-to-end

optimization and better performance on face detection.

3.1.2 Single-stage methods. The single-stage methods accomplish the candidate classication and

bounding box regression from the entire feature maps directly, without involving the proposal

stage.

A classic structure of single stage comes from a general object detector named Single Shot

multibox Detector (SSD) [

142

]. Similar to RPN, SSD presets dense anchor boxes over dierent

ratios and scales on the feature maps. SSD is a prevailing framework in object detection because it

runs much faster than Faster R-CNN while maintaining comparable accuracy. So, many developers

employed SSD for face detection in applications. However, SSD is not robust enough to large scale

variation, especially to the small faces. Afterward, many methods [

224

327

–

329

] studied to

modify SSD for face detection. For example, Zhang et al. [

328

] designed a scale-equitable version

to obtain adequate features from the faces of dierent scales.

Many state-of-the-art face detectors resort to the feature pyramid network (FPN) [

132

] which

consists of a top-down architecture with skip connections and merge the high-level and low-level

features for detection. The high-level feature maps have more semantic information, while the

low-level layers have smaller receptive eld but more detailed local information. The feature fusion

preserves the advantages from both sides, and brings great progress in detecting objects with a

wide range of scales. Therefore, many single-stage face detectors [

124

130

168

224

225

244

318

326

] are developed with the advantage of FPN. Not only handling the scale issue in face

detection via FPN, but also these methods attempt to solve the inherent shortcomings of original

FPN such like the conict of receptive eld. The special feature fusion operation [

124

130

224

] is

also suitable for tackling the hard cases of face detection, such as blur and occluded faces.

Although the single-stage methods have the advantage of high eciency, their detection accuracy

is below that of the two-stage methods. It is partially because the imbalance problem of positives

and negatives brought by the dense anchors, whereas the proposal-to-rene scheme is able to

alleviate this issue. Accordingly, ReneDet [

325

] set up an anchor renement module in its network

to remove large number of negatives. Inspired by ReneDet, SRN [

] presented a selective two-step

classication and regression method; the two-step classication is performed at the low-level layers

to reduce the search space of classier, and the two-step regression is performed at high-level

layers to obtain accurate location. Later on, VIM-FD [

337

], ISRN [

326

], AInnoFace [

313

] and Re-

neFace [

323

] improved SRN with several eective techniques, such as training data augmentation,

improved feature extractor and training supervision, anchor assignment and matching strategy,

multi-scale test strategy etc.

Most aforementioned methods need to preset anchors for face detection, while some representa-

tive detectors of single-stage, such as DenseBox [

], UnitBox [

298

] and CenterFace [

280

], full the

detection without preset anchors. We will present them as anchor-free type in the next subsection.

3.1.3 Anchor-based and anchor-free methods. As shown in Table 2, most current face detectors are

anchor-based due to the long-time development and superior performance. Generally, we preset

the dense anchors on the feature maps, then full the classication and bounding box regression

on these anchors one or more times, and nally output the accepted ones as the detection results.

Therefore, the anchor allocation and matching strategy is crucial to the detection accuracy. For

example, the scale compensation for anchor matching, proposed by S

FD [

328

], can eectively

improve the recall of tiny and outer faces. Besides, S

FD utilized a max-out label mechanism to

reduce the large number of negatives which is a frequent issue in anchor-based mechanism as

well. Zhu et al. [

356

] introduced an expected max overlapping score (EMO) to evaluate the quality

of matched anchors, and proposed several techniques to encourage the true positives achieve

high EMO scores. Since the scale distribution of faces is imbalance in the training dataset, Group

Sampling [

164

] sorts the anchor boxes by their scales and maintains the same number of samples

for each group during the training. More recently, HAMBox [

151

] proposed an online anchor

compensation strategy to help the detection of outer faces, taking the advantage of unmatched

anchors that nonetheless provide favorable regression.

The anchor-based methods have dominated state of the art in face detection, but there are several

weaknesses of them. The hyperparameters ( e.g., scale, stride, ratio, number) of preset anchors need

to be carefully tuned for each particular dataset, which limits the generalization ability of detectors.

Besides, the dense anchors increase the computational cost and bring the imbalance problem of

positive and negative anchors.

Anchor-free methods [

120

226

355

] attract growing attention in general object detection. As

for face detection, certain pioneering works have emerged in recent years. DenseBox [

] and

UnitBox [

298

] attempt to predict the pixel-wise bounding box and the condence score. Besides,

CenterFace [

280

] regards face detection as a generalized task of keypoint estimation, which predicts

the facial center point and the size of bounding box in feature map. In brief, the anchor-free

detectors get rid of the preset anchors and achieve the better generalization capacity. Regarding

to the detection accuracy, it needs further exploration for better robustness to false positives and

stability in training process.

3.1.4 Multi-task learning methods. Multi-task learning has been widely studied in computer vision

community. Generally, the multi-task learning based approaches are designed for solving a problem

together with other related tasks by sharing the visual representation. Here, we introduce the

multi-task learning methods that trains the face detector with the associated facial tasks or auxiliary

supervision branches to enrich the feature representation and detection robustness.

Many multi-task learning methods [

128

280

310

319

367

] have explored the joint

learning of face detection and facial landmark localization. Among them, MTCNN [

319

] is the

most representative one, which exploits the inherent correlation between facial bounding boxes

and landmarks by a three-stage cascaded network. Subsequently, HyperFace [

180

] fused the low-

level features as well as the high-level features to simultaneously conduct four tasks, including

face detection, facial landmark localization, gender classication and pose estimation. Based on

剩余43页未读，继续阅读

puluowangsi2

粉丝: 0
资源: 19

"深度人脸识别综述：最新进展与关键元素"

人脸识别 论文

论文研究-人脸识别方法综述.pdf

论文研究-人脸识别研究综述.pdf

人脸识别演示界面：python GUI--tkinter代码

python人脸识别操作界面_人脸识别演示界面：python GUI--tkinter实战（2）

lbp 人脸识别 matlab,GitHub - stuhard/LBP-DBN-face-recognition: 使用LBP特征提取算法提取人脸特征，DBN网络来实现人脸识别，测试数据库ORL数据...

Combining 3D Morphable Models: A Large scale Face-and-Head Model

类似opencv人脸识别的项目python

列举20个ml.net框架下第三方开源项目，关于人脸识别的。并给出下载地址

最新资源

人脸识别论文