专利快速检索-快速检索全球专利，免费商用专利数据库-IPRDB

1. 发明专利

GB2582833B Facial localisation in images 未知
公开(公告)号：GB2582833B
公开(公告)日：2021-04-07
申请号：GB201906027
申请日：2019-04-30
申请人： HUAWEI TECH CO LTD
发明人： JIANKANG DENG , STEFANOS ZAFEIRIOU
IPC分类号： G06K9/00 , G06N3/04 , G06N3/08
摘要： The invention relates to methods of facial localisation in images. According to a first aspect (Fig. 2), there is described a computer-implemented method of training a neural network 106 for face localisation, the method comprising: inputting, to the neural network, a training image 202 comprising one or more faces (104); processing, using the neural network, the training image and outputting for each of a plurality of training anchors in the training image, one or more sets of output data, wherein the output data comprises a predicted facial classification, a predicted location of a corresponding face box (110), and one or more corresponding feature vectors; then updating parameters of the neural network in dependence on an objective function. The objective function comprises, for each positive anchor in the training image, a classification loss 210, a box regression loss 212 and a feature loss 214, based on comparisons of predicted properties with known properties in the anchor. A second aspect (Fig. 1) covers using a neural network to perform face localisation, wherein the network comprises convolutional layers and filters, as well as skip connections.

2. 发明专利

GB2592076A Method of training an image classification model 未知
公开(公告)号：GB2592076A
公开(公告)日：2021-08-18
申请号：GB202002157
申请日：2020-02-17
申请人： HUAWEI TECH CO LTD
发明人： JIANKANG DENG , STEFANOS ZAFEIRIOU
IPC分类号： G06K9/62 , G06N3/08
摘要： Method of training a neural network classifier, comprising: extracting from the neural network 204 a plurality of subclass centre vectors 210; generating an embedding vector 206 from an input image 202; determining a similarity score 212 between the embedding vector and each of the subclass centre vectors; updating neural network parameters 214 in dependence each of the similarity scores using an objective function; and extracting and updating each subclass centre vector from the neural network. Subclass centre vectors may be extracted from the neural network’s last fully connected layer. The objective function may comprise a multi-centre loss term such as a margin-based Softmax loss function, that determines, for the embedding vector, a closest subclass centre vector for each class using the similarity scores. The embedding vector and subclass centre vectors may be normalised, and the similarity score an angle therebetween. The objective function may comprise an intra-class compactness term that uses the intra-class normalized angle similarity score between sub-class a dominant vector and the other sub-classes. Non-dominant subclasses may be discarded after a number of training cycles have been run. The method may be used to train a neural network on face images that contain noise (wrong labels).

3. 发明专利

GB2576784B Facial landmark localisation system and method 未知
公开(公告)号：GB2576784B
公开(公告)日：2021-05-19
申请号：GB201814275
申请日：2018-09-03
申请人： HUAWEI TECH CO LTD
发明人： JIANKANG DENG , STEFANOS ZAFEIRIOU
IPC分类号： G06K9/46 , G06N3/04
摘要： A neural network system and method for localising facial landmarks in an image comprises two or more convolutional neural networks, 200, in series. Each convolution neural network comprises a plurality of downsampling layers, 211, 212, 213, 224, which downsample the input image signal, and a plurality of upsampling layers 220, 222, 216, 224, 215, 214 which upsample the downsampled signal. During the upsampling of the downsampled signal at an upsampling layer, the upsampled signal is also combined with the signal from a connection to a lateral layer of equal size. Each of the upsampling layers thus aggregates input from the previous lateral layer of equal size with downsampled input from a smaller downsampling layer. At least one of these upsampling layers, 224, further includes an input from a larger (upsampling) layer. Lateral connections may be skip connections or one or more convolutions, e.g. depth-wise separable convolution(s). The input may be a 128x128 pixel input image. The system may comprise a channel aggregation block for each layer of the convolutional neural network, blocks including channel increase, decrease and branch steps. Outputs of each convolutional neural network may be connected to spatial transformers, such as a deformable convolution.

4. 发明专利

GB2582833A Facial localisation in images 未知
公开(公告)号：GB2582833A
公开(公告)日：2020-10-07
申请号：GB201906027
申请日：2019-04-30
申请人： HUAWEI TECH CO LTD
发明人： JIANKANG DENG , STEFANOS ZAFEIRIOU
IPC分类号： G06K9/00 , G06N3/04 , G06N3/08
摘要： The invention relates to methods of facial localisation in images. According to a first aspect (Fig. 2), there is described a computer-implemented method of training a neural network 106 for face localisation, the method comprising: inputting, to the neural network, a training image 202 comprising one or more faces (104); processing, using the neural network, the training image and outputting for each of a plurality of training anchors in the training image, one or more sets of output data, wherein the output data comprises a predicted facial classification, a predicted location of a corresponding face box (110), and one or more corresponding feature vectors; then updating parameters of the neural network in dependence on an objective function. The objective function comprises, for each positive anchor in the training image, a classification loss 210, a box regression loss 212 and a feature loss 214, based on comparisons of predicted properties with known properties in the anchor. A second aspect (Fig. 1) covers using a neural network to perform face localisation, wherein the network comprises convolutional layers and filters, as well as skip connections.

5. 发明专利

GB2592076B Method of training an image classification model 未知
公开(公告)号：GB2592076B
公开(公告)日：2022-09-07
申请号：GB202002157
申请日：2020-02-17
申请人： HUAWEI TECH CO LTD
发明人： JIANKANG DENG , STEFANOS ZAFEIRIOU
IPC分类号： G06V10/77 , G06K9/62 , G06N3/08
摘要： Method of training a neural network classifier, comprising: extracting from the neural network 204 a plurality of subclass centre vectors 210; generating an embedding vector 206 from an input image 202; determining a similarity score 212 between the embedding vector and each of the subclass centre vectors; updating neural network parameters 214 in dependence each of the similarity scores using an objective function; and extracting and updating each subclass centre vector from the neural network. Subclass centre vectors may be extracted from the neural network’s last fully connected layer. The objective function may comprise a multi-centre loss term such as a margin-based Softmax loss function, that determines, for the embedding vector, a closest subclass centre vector for each class using the similarity scores. The embedding vector and subclass centre vectors may be normalised, and the similarity score an angle therebetween. The objective function may comprise an intra-class compactness term that uses the intra-class normalized angle similarity score between sub-class a dominant vector and the other sub-classes. Non-dominant subclasses may be discarded after a number of training cycles have been run. The method may be used to train a neural network on face images that contain noise (wrong labels).

你已经成功收藏专利！

检索式保存成功!

IPRDB

热门服务

关于我们

友情链接

联系方式