WO2020233368A1 - 表情识别模型训练方法、装置、设备和存储介质 - Google Patents
表情识别模型训练方法、装置、设备和存储介质 Download PDFInfo
- Publication number
- WO2020233368A1 WO2020233368A1 PCT/CN2020/087605 CN2020087605W WO2020233368A1 WO 2020233368 A1 WO2020233368 A1 WO 2020233368A1 CN 2020087605 W CN2020087605 W CN 2020087605W WO 2020233368 A1 WO2020233368 A1 WO 2020233368A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- image
- sub
- original
- training
- resolution
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Ceased
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
- G06V40/174—Facial expression recognition
Definitions
- This application relates to the field of image processing technology, in particular to an expression recognition model training method, device, equipment and storage medium.
- the facial expression recognition model is used to recognize facial expressions. Facial expression recognition refers to assigning an expression category to a given facial image, including: anger, disgust, happiness, sadness, fear, surprise, etc.
- the inventor realizes that the current facial expression recognition technology is gradually showing broad application prospects in the fields of human-computer interaction, clinical diagnosis, remote education, and investigation and interrogation, and it is a popular research direction in computer vision and artificial intelligence.
- the facial expression recognition model needs to be trained in advance.
- the existing expression recognition model training method uses The resolution or tones of the training images are the same or similar, resulting in the expression recognition model after the training can more accurately recognize the expression image within a fixed resolution or tonal range, and the resolution or tone of the same expression image is reduced Changes will reduce the recognition accuracy of the facial expression recognition model.
- the main purpose of this application is to solve the technical problem that the existing expression recognition model training method is time-consuming and laborious to mark training images, and the recognition accuracy of the trained expression recognition model is easily affected by the resolution and tone of the expression image.
- An expression recognition model training method including: acquiring an original training image set; the original training image set includes a plurality of labeled original training images; and the following processing is performed on the original training image set: reducing the original training image The resolution of each original training image in the set to obtain the first type of training image set; render the background light of each original training image in the original training image set to obtain the second type of training image set; reduce the original training image The resolution of each original training image in the set, and render the background light of each original training image to obtain the third type of training image set; respectively pass the original training image set, the first type training image set, and the first type The second-type training image set and the third-type training image set train the facial expression recognition model.
- an expression recognition model training device which includes: an acquisition module for acquiring the original training image set.
- the original training image set includes a plurality of labeled original training images.
- the processing module is configured to perform the following processing on the original training image set obtained by the obtaining module: reduce the resolution of each original training image in the original training image set to obtain the first type of training image set; rendering The background light of each original training image in the original training image set to obtain the second type of training image set; reduce the resolution of each original training image in the original training image set, and render the background light of each original training image , Get the third type of training image set.
- the processing module is further configured to train an expression recognition model through the original training image collection, the first type training image collection, the second type training image collection, and the third type training image collection respectively.
- the present application also provides a computer device, including an input and output unit, a memory, and a processor.
- the memory stores computer-readable instructions that are executed by the processor. , Making the processor execute the steps in the above-mentioned expression recognition model training method.
- the present application also provides a storage medium storing computer-readable instructions.
- the computer-readable instructions are executed by one or more processors, one or more processors can execute the above Steps in the training method of facial expression recognition model.
- This application obtains multiple types of new training images by adjusting the features such as the clarity or background tone of the original training images.
- the new training images do not need to be manually marked, which enriches the training sample image set of the facial expression recognition model.
- FIG. 1 is a schematic flowchart of a method for training an expression recognition model in an embodiment of the application.
- Fig. 2 is a schematic structural diagram of an expression recognition model training device in an embodiment of the application.
- Fig. 3 is a schematic structural diagram of a computer device in an embodiment of the application.
- FIG. 1 is a flowchart of an expression recognition model training method in some embodiments of the application.
- the expression recognition model training method is executed by an expression recognition model training device.
- the expression recognition model training device may be a computer or other devices, as shown in FIG. 1 , Can include the following steps S1-S3:
- the original training image set includes a plurality of labeled original training images.
- the original training image is a manually labeled training sample image, which is used to train the facial expression recognition model.
- the number of training sample images required for expression recognition model training is very large.
- the traditional method of marking training sample images is to manually label training sample images one by one, which consumes a lot of time and labor costs.
- a deep neural network model is used to reduce the resolution of each original training image in the original training image set.
- the deep neural network model is generated by the model generation device according to low-resolution image samples, image conversion algorithms, and a deep neural network framework.
- the deep neural network model includes a plurality of nonlinear conversion convolutional layers alternately using different parameter matrices as convolution template parameters.
- step S1 the method further includes the following steps S01-S03:
- Segment low-resolution image samples to enrich the set of low-resolution image samples.
- S02. Perform image conversion on the low-resolution sub-image samples by using an image conversion algorithm to obtain high-resolution sub-image samples corresponding to the low-resolution sub-image samples.
- step S02 includes the following steps S021-S024:
- Total Variation also known as Total Variation
- Total Variation is often used for image restoration.
- Image decomposition is used to decompose low-resolution sub-image samples into cartoon and texture parts.
- the cartoon part extracts the structural information of the low-resolution sub-image samples.
- the pixel value only changes greatly at the boundary of the object, and the pixel value inside the object changes little, and the image is smooth.
- the texture part extracts the detailed part of the low-resolution sub-image sample, in which the pixel value changes greatly.
- the expression of the image total variation algorithm is:
- (x p , y p ) represents the current central pixel in the low-resolution sub-image sample;
- (x q , y q ) represents the total variational pixel of (x p , y p );
- T g is the preset threshold;
- S022 Use an interpolation algorithm to enlarge the cartoon sub-image sample to obtain an enlarged cartoon sub-image sample.
- the pixel points of the cartoon sub-image sample are interpolated using an interpolation template function to obtain the enlarged cartoon sub-image sample.
- Image interpolation belongs to the prior art, and will not be repeated here.
- Homotopy Method describes the “continuous change” between two objects in topology. If two topological spaces can be changed from one to another through a series of continuous deformations, then the two topological spaces are said to be the same. Lun.
- step S023 includes the following steps: using a dictionary training algorithm to obtain an image block dictionary of the texture sub-image sample; using the image block dictionary and an orthogonal matching tracking method to amplify the texture sub-image sample to obtain An initial high-resolution sub-image; perform nearest neighbor edge addition processing to the initial high-resolution sub-image to obtain an edged high-resolution sub-image; perform the first homotopy processing on the edged high-resolution sub-image to obtain A first edged high-resolution sub-image; a second homotopic processing is performed on the first edged high-resolution sub-image to obtain the enlarged texture sub-image sample.
- the image block dictionary includes a high-resolution image block dictionary and a low-resolution image block dictionary.
- the dictionary training algorithm is a K-SVD dictionary training algorithm.
- K-SAD is a classic dictionary training algorithm.
- the error term is subjected to Singular Value Decomposition (SVD), and the decomposition term that minimizes the error is selected as the updated dictionary atom and corresponding atomic coefficient. After continuous iteration, an optimized solution is obtained.
- SVD Singular Value Decomposition
- the expression of image synthesis is:
- f H is the high-resolution sub-image sample
- f t is the enlarged texture image sample
- f c is the enlarged cartoon image sample
- G(f t ) is the modulus of the Robert gradient of the image f t
- ⁇ 1 is a constant greater than zero.
- the high-resolution sub-image sample is an image after resolution conversion of the low-resolution sub-image sample.
- the background light rendering of the image is a process of toning the background of the image, which belongs to the prior art and will not be repeated here.
- the facial expression recognition model is used to recognize the micro-expressions of people in facial images, such as happy, sad, fearful, angry, surprised and disgusted.
- the obtained training images of the first type, the second type, and the third type of training images are new images for the expression recognition model. Therefore, based on the original training image that has been annotated, the obtained first type training image collection, second type training image collection, and third type training image collection do not need to be manually marked, and the training samples of the expression recognition model are enriched Image set.
- step S3 the method further includes the following steps S4-S7:
- the original test image set includes a plurality of original test images.
- the original test image is used to test the accuracy of facial image recognition by the trained expression recognition model.
- S5. Perform the following processing on the original test image set: reduce the resolution of each original test image in the original test image set to obtain the first type of test image set; render each original test image in the original test image set Test the background light of the image to obtain the second type of test image set; reduce the resolution of each original test image in the original test image set, and render the background light of each original test image to obtain the third type of test image set.
- the resolution of the original test image and the processing process of the background light are the same as the resolution of the original training image and the processing process of the background light, which will not be repeated here.
- the trained expression recognition model recognizes the first-type test image set, and outputs the recognition result of each first-type test image; compares the recognition result with a preset The results are compared, and if the recognition result is consistent with the preset comparison result, it is determined that the recognition result output by the expression recognition model is correct; otherwise, it is determined that the recognition result output by the expression recognition model is wrong. Record the number of accurately identified test images of the first type, and divide the number of accurately identified test images of the first type by the total number of test images of the first type to obtain the expression recognition model for the first type of test images Accuracy of collection recognition.
- the characteristics of the original training image such as the sharpness or background hue
- multiple types of new training images are obtained.
- the new training images do not need to be manually marked, which enriches the training sample images of the expression recognition model Collection, greatly reducing the time and labor cost of training sample image marking operations.
- statistics of the recognition accuracy of the facial expression recognition model on the original training image, the first type of training image, the second type of training image, and the third type of training image are provided for evaluating the actual effect of the facial expression recognition model in accordance with.
- this application also provides an expression recognition model training device, which can be used to enrich the training image set and improve the efficiency of expression recognition model training.
- the device in the embodiment of the present application can implement the steps corresponding to the method for training an expression recognition model performed in the embodiment corresponding to FIG. 1.
- the functions realized by the device can be realized by hardware, or by hardware executing corresponding software.
- the hardware or software includes one or more modules corresponding to the above functions, and the modules may be software and/or hardware.
- the device includes an acquisition module 1 and a processing module 2.
- the processing module 2 can be used to control the receiving and sending operations of the acquiring module 1.
- the acquisition module 1 is used to acquire the original training image set.
- the original training image set includes a plurality of labeled original training images.
- the processing module 2 is configured to perform the following processing on the acquired original training image set acquired by the acquisition module 1 respectively: reduce the resolution of each original training image in the original training image set to obtain the first One type of training image set; rendering the background light of each original training image in the original training image set to obtain the second type of training image set; reducing the resolution of each original training image in the original training image set, and rendering The background light of each original training image obtains the third type of training image set.
- the processing module 2 is also configured to train an expression recognition model through the original training image collection, the first type training image collection, the second type training image collection, and the third type training image collection.
- the acquisition module 1 is also used to acquire an original test image set; the original test image set includes a plurality of original test images; the original test images are used to test the recognition of facial images by the expression recognition model after training. Accuracy.
- the processing module 2 is also configured to perform the following processing on the original test image set acquired by the acquisition module 1 respectively: reduce the resolution of each original test image in the original test image set to obtain the first type of test Image collection; rendering the background light of each original test image in the original test image collection to obtain the second type of test image collection; reducing the resolution of each original test image in the original test image collection, and rendering each original test
- reduce the resolution of each original test image in the original test image set is obtained; the original test image set, the first type test image set, the second type test image set, and the third type Class test image collection for identification.
- the processing module 2 is also used to separately count the training of the facial expression recognition model on the original test image set, the first type test image set, the second type test image set, and the third type test image set. The accuracy of image collection recognition.
- a deep neural network model is used to reduce the resolution of each original training image in the original training image set.
- the processing module 2 is further configured to use high-resolution sub-image samples as input samples of the deep neural network framework, and use low-resolution sub-image samples as output comparison samples of the deep neural network framework to generate the deep neural network model ;
- the high-resolution sub-image sample is the image after the resolution conversion of the low-resolution sub-image sample.
- the processing module 2 is further used to divide the low-resolution image sample into multiple low-resolution sub-image samples; image conversion is performed on the low-resolution sub-image samples using an image conversion algorithm to obtain low-resolution sub-image samples.
- the high-resolution sub-image sample corresponding to the image sample.
- the processing module 2 is specifically configured to decompose low-resolution sub-image samples using an image total variation algorithm to obtain cartoon sub-image samples and texture sub-image samples; use an interpolation algorithm to analyze the cartoon sub-image samples Zoom in to obtain an enlarged cartoon sub-image sample; use the homotopy method to enlarge the texture sub-image sample to obtain an enlarged texture sub-image sample; compare the enlarged cartoon sub-image sample and the enlarged
- the texture sub-image samples are synthesized to obtain high-resolution sub-image samples.
- the expression of the image total variation algorithm is:
- (x p , y p ) represents the current central pixel in the low-resolution sub-image sample;
- (x q , y q ) represents the total variational pixel of (x p , y p );
- T g is the preset threshold;
- the processing module 2 is specifically configured to use a dictionary training algorithm to obtain an image block dictionary of the texture sub-image sample; use the image block dictionary and an orthogonal matching tracking method to amplify the texture sub-image sample , Obtain an initial high-resolution sub-image; perform nearest neighbor edge addition processing to the initial high-resolution sub-image to obtain an edge-added high-resolution sub-image; perform the first homotopy processing on the edge-added high-resolution sub-image To obtain a first edged high-resolution sub-image; perform a second homotopy process on the first edged high-resolution sub-image to obtain the enlarged texture sub-image sample.
- the characteristics of the original training image such as the sharpness or background hue
- multiple types of new training images are obtained.
- the new training images do not need to be manually marked, which enriches the training sample images of the expression recognition model Collection, greatly reducing the time and labor cost of training sample image marking operations.
- statistics of the recognition accuracy of the facial expression recognition model on the original training image, the first type of training image, the second type of training image, and the third type of training image are provided for evaluating the actual effect of the facial expression recognition model in accordance with.
- the present application also provides a computer device, as shown in FIG. 3, the computer device includes an input output unit 31, a processor 32, and a memory 33.
- the memory 33 stores computer readable instructions, When the computer-readable instructions are executed by the processor 32, the processor executes the steps of the expression recognition model training method in the foregoing embodiments.
- the physical device corresponding to the acquisition module 1 shown in FIG. 2 is the input and output unit 31 shown in FIG. 3, which can realize part or all of the functions of the acquisition module 1, or realize the same or similar functions as the acquisition module 1.
- the physical device corresponding to the processing module 2 shown in FIG. 2 is the processor 32 shown in FIG. 3, and the processor 32 can implement part or all of the functions of the processing module 2 or implement the same or similar functions as the processing module 2.
- the present application also provides a storage medium storing computer-readable instructions.
- the computer-readable storage medium may be non-volatile or volatile.
- the one or more processors execute the steps of the expression recognition model training method in the foregoing embodiments.
- the method of the above embodiments can be implemented by means of software plus the necessary general hardware platform. Of course, it can also be implemented by hardware, but in many cases the former is better. ⁇
- the technical solution of this application essentially or the part that contributes to the existing technology can be embodied in the form of a software product.
- the computer software product is stored in a storage medium (such as ROM/RAM), including Several instructions are used to make a terminal (which may be a mobile phone, a computer, a server, or a network device, etc.) execute the method described in each embodiment of the present application.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Oral & Maxillofacial Surgery (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
Description
Claims (21)
- 一种表情识别模型训练方法,其中,包括:获取原训练图像集合;所述原训练图像集合包括多个已标注的原训练图像;对所述原训练图像集合分别进行以下处理:降低所述原训练图像集合中的各原训练图像的分辨率,得到第一类训练图像集合;渲染所述原训练图像集合中的各原训练图像的背景光线,得到第二类训练图像集合;降低所述原训练图像集合中的各原训练图像的分辨率,并且渲染各原训练图像的背景光线,得到第三类训练图像集合;分别通过所述原训练图像集合、所述第一类训练图像集合、所述第二类训练图像集合以及所述第三类训练图像集合训练表情识别模型。
- 根据权利要求1所述的表情识别模型训练方法,其中,在所述分别通过所述原训练图像集合、所述第一类训练图像集合、所述第二类训练图像集合以及所述第三类训练图像集合训练表情识别模型之后,所述方法还包括:获取原测试图像集合;所述原测试图像集合包括多个原测试图像;原测试图像用于测试训练后的表情识别模型对人脸图像识别的准确率;对所述原测试图像集合分别进行以下处理:降低所述原测试图像集合中的各原测试图像的分辨率,得到第一类测试图像集合;渲染所述原测试图像集合中的各原测试图像的背景光线,得到第二类测试图像集合;降低所述原测试图像集合中的各原测试图像的分辨率,并且渲染各原测试图像的背景光线,得到第三类测试图像集合;通过训练后的表情识别模型对所述原测试图像集合、所述第一类测试图像集合、所述第二类测试图像集合以及所述第三类测试图像集合进行识别;分别统计所述训练后的表情识别模型对所述原测试图像集合、所述第一类测试图像集合、所述第二类测试图像集合以及所述第三类测试图像集合识别的准确率。
- 根据权利要求1所述的表情识别模型训练方法,其中,采用深度神经网络模型降低所述原训练图像集合中的各原训练图像的分辨率;在所述获取原训练图像集合之前,所述方法还包括:以高分辨率子图像样本作为深度神经网络框架的输入样本,以低分辨率子图像样本作为所述深度神经网络框架的输出对比样本,生成所述深度神经网络模型;高分辨率子图像样本为低分辨率子图像样本分辨率转化后的图像。
- 根据权利要求3所述的表情识别模型训练方法,其中,在所述以高分辨率子图像样本作为深度神经网络框架的输入样本,以低分辨率子图像样本作为所述深度神经网络框架的输出对比样本,生成所述深度神经网络模型之前,所述方法还包括:将低分辨率图像样本分割为多个低分辨率子图像样本;采用图像转换算法对低分辨率子图像样本进行图像转换,得到低分辨率子图像样本对应的高分辨率子图像样本。
- 根据权利要求4所述的表情识别模型训练方法,其中,所述采用图像转换算法对低分辨率子图像样本进行图像转换,得到低分辨率子图像样本对应的高分辨率子图像样本,包括:采用图像全变分算法对低分辨率子图像样本进行分解,得到卡通子图像样本和纹理子图像样本;采用插值算法对所述卡通子图像样本进行放大,得到放大后的卡通子图像样本;采用同伦法对所述纹理子图像样本进行放大,得到放大后的纹理子图像样本;对所述放大后的卡通子图像样本和所述放大后的纹理子图像样本进行合成,得到高分辨率子图像样本。
- 根据权利要求5所述的表情识别模型训练方法,其中,所述采用同伦法对所述纹理子图像样本进行放大,得到放大后的纹理子图像样本,包括:采用字典训练算法得到所述纹理子图像样本的图像块字典;采用所述图像块字典和正交匹配跟踪方法对所述纹理子图像样本进行放大,得到初始高分辨率子图像;对所述初始高分辨率子图像进行最近邻的加边处理,得到加边高分辨子图像;对所述加边高分辨子图像进行第一次同伦处理,得到第一加边高分辨率子图像;对所述第一加边高分辨率子图像进行第二次同伦处理,得到所述放大后的纹理子图像样本。
- 一种计算机设备,其中,包括输入输出单元、存储器和处理器,所述存储器中存储有计算机可读指令,所述计算机可读指令被所述处理器执行时,使得所述处理器执行如下的步骤:获取原训练图像集合;所述原训练图像集合包括多个已标注的原训练图像;对所述原训练图像集合分别进行以下处理:降低所述原训练图像集合中的各原训练图像的分辨率,得到第一类训练图像集合;渲染所述原训练图像集合中的各原训练图像的背景光线,得到第二类训练图像集合;降低所述原训练图像集合中的各原训练图像的分辨率,并且渲染各原训练图像的背景光线,得到第三类训练图像集合;分别通过所述原训练图像集合、所述第一类训练图像集合、所述第二类训练图像集合以及所述第三类训练图像集合训练表情识别模型。
- 根据权利要求8所述的计算机设备,其中,在所述分别通过所述原训练图像集合、所述第一类训练图像集合、所述第二类训练图像集合以及所述第三类训练图像集合训练表情识别模型之后,所述方法还包括:获取原测试图像集合;所述原测试图像集合包括多个原测试图像;原测试图像用于测试训练后的表情识别模型对人脸图像识别的准确率;对所述原测试图像集合分别进行以下处理:降低所述原测试图像集合中的各原测试图像的分辨率,得到第一类测试图像集合;渲染所述原测试图像集合中的各原测试图像的背景光线,得到第二类测试图像集合;降低所述原测试图像集合中的各原测试图像的分辨率,并且渲染各原测试图像的背景光线,得到第三类测试图像集合;通过训练后的表情识别模型对所述原测试图像集合、所述第一类测试图像集合、所述第二类测试图像集合以及所述第三类测试图像集合进行识别;分别统计所述训练后的表情识别模型对所述原测试图像集合、所述第一类测试图像集合、所述第二类测试图像集合以及所述第三类测试图像集合识别的准确率。
- 根据权利要求8所述的计算机设备,其中,采用深度神经网络模型降低所述原训练图像集合中的各原训练图像的分辨率;在所述获取原训练图像集合之前,所述方法还包括:以高分辨率子图像样本作为深度神经网络框架的输入样本,以低分辨率子图像样本作为所述深度神经网络框架的输出对比样本,生成所述深度神经网络模型;高分辨率子图像样本为低分辨率子图像样本分辨率转化后的图像。
- 根据权利要求10所述的计算机设备,其中,在所述以高分辨率子图像样本作为深度神经网络框架的输入样本,以低分辨率子图像样本作为所述深度神经网络框架的输出对比样本,生成所述深度神经网络模型之前,所述方法还包括:将低分辨率图像样本分割为多个低分辨率子图像样本;采用图像转换算法对低分辨率子图像样本进行图像转换,得到低分辨率子图像样本对应的高分辨率子图像样本。
- 根据权利要求11所述的计算机设备,其中,所述采用图像转换算法对 低分辨率子图像样本进行图像转换,得到低分辨率子图像样本对应的高分辨率子图像样本,包括:采用图像全变分算法对低分辨率子图像样本进行分解,得到卡通子图像样本和纹理子图像样本;采用插值算法对所述卡通子图像样本进行放大,得到放大后的卡通子图像样本;采用同伦法对所述纹理子图像样本进行放大,得到放大后的纹理子图像样本;对所述放大后的卡通子图像样本和所述放大后的纹理子图像样本进行合成,得到高分辨率子图像样本。
- 根据权利要求13所述的计算机设备,其中,所述采用同伦法对所述纹理子图像样本进行放大,得到放大后的纹理子图像样本,包括:采用字典训练算法得到所述纹理子图像样本的图像块字典;采用所述图像块字典和正交匹配跟踪方法对所述纹理子图像样本进行放大,得到初始高分辨率子图像;对所述初始高分辨率子图像进行最近邻的加边处理,得到加边高分辨子图像;对所述加边高分辨子图像进行第一次同伦处理,得到第一加边高分辨率子图像;对所述第一加边高分辨率子图像进行第二次同伦处理,得到所述放大后的纹理子图像样本。
- 一种存储有计算机可读指令的存储介质,其中,所述计算机可读指令被一个或多个处理器执行时,使得一个或多个处理器执行如下的步骤:获取原训练图像集合;所述原训练图像集合包括多个已标注的原训练图像;对所述原训练图像集合分别进行以下处理:降低所述原训练图像集合中的各原训练图像的分辨率,得到第一类训练图像集合;渲染所述原训练图像集合中的各原训练图像的背景光线,得到第二类训练图像集合;降低所述原训练图像集合中的各原训练图像的分辨率,并且渲染各原训练 图像的背景光线,得到第三类训练图像集合;分别通过所述原训练图像集合、所述第一类训练图像集合、所述第二类训练图像集合以及所述第三类训练图像集合训练表情识别模型。
- 根据权利要求15所述的存储介质,其中,在所述分别通过所述原训练图像集合、所述第一类训练图像集合、所述第二类训练图像集合以及所述第三类训练图像集合训练表情识别模型之后,所述方法还包括:获取原测试图像集合;所述原测试图像集合包括多个原测试图像;原测试图像用于测试训练后的表情识别模型对人脸图像识别的准确率;对所述原测试图像集合分别进行以下处理:降低所述原测试图像集合中的各原测试图像的分辨率,得到第一类测试图像集合;渲染所述原测试图像集合中的各原测试图像的背景光线,得到第二类测试图像集合;降低所述原测试图像集合中的各原测试图像的分辨率,并且渲染各原测试图像的背景光线,得到第三类测试图像集合;通过训练后的表情识别模型对所述原测试图像集合、所述第一类测试图像集合、所述第二类测试图像集合以及所述第三类测试图像集合进行识别;分别统计所述训练后的表情识别模型对所述原测试图像集合、所述第一类测试图像集合、所述第二类测试图像集合以及所述第三类测试图像集合识别的准确率。
- 根据权利要求15所述的存储介质,其中,采用深度神经网络模型降低所述原训练图像集合中的各原训练图像的分辨率;在所述获取原训练图像集合之前,所述方法还包括:以高分辨率子图像样本作为深度神经网络框架的输入样本,以低分辨率子图像样本作为所述深度神经网络框架的输出对比样本,生成所述深度神经网络模型;高分辨率子图像样本为低分辨率子图像样本分辨率转化后的图像。
- 根据权利要求17所述的存储介质,其中,在所述以高分辨率子图像样本作为深度神经网络框架的输入样本,以低分辨率子图像样本作为所述深度神经网络框架的输出对比样本,生成所述深度神经网络模型之前,所述方法还包括:将低分辨率图像样本分割为多个低分辨率子图像样本;采用图像转换算法对低分辨率子图像样本进行图像转换,得到低分辨率子图像样本对应的高分辨率子图像样本。
- 根据权利要求18所述的存储介质,其中,所述采用图像转换算法对低分辨率子图像样本进行图像转换,得到低分辨率子图像样本对应的高分辨率子图像样本,包括:采用图像全变分算法对低分辨率子图像样本进行分解,得到卡通子图像样本和纹理子图像样本;采用插值算法对所述卡通子图像样本进行放大,得到放大后的卡通子图像样本;采用同伦法对所述纹理子图像样本进行放大,得到放大后的纹理子图像样本;对所述放大后的卡通子图像样本和所述放大后的纹理子图像样本进行合成,得到高分辨率子图像样本。
- 根据权利要求20所述的存储介质,其中,所述采用同伦法对所述纹理子图像样本进行放大,得到放大后的纹理子图像样本,包括:采用字典训练算法得到所述纹理子图像样本的图像块字典;采用所述图像块字典和正交匹配跟踪方法对所述纹理子图像样本进行放大,得到初始高分辨率子图像;对所述初始高分辨率子图像进行最近邻的加边处理,得到加边高分辨子图像;对所述加边高分辨子图像进行第一次同伦处理,得到第一加边高分辨率子图像;对所述第一加边高分辨率子图像进行第二次同伦处理,得到所述放大后的纹理子图像样本。
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910427443.1 | 2019-05-22 | ||
| CN201910427443.1A CN110309713A (zh) | 2019-05-22 | 2019-05-22 | 表情识别模型训练方法、装置、设备和存储介质 |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| WO2020233368A1 true WO2020233368A1 (zh) | 2020-11-26 |
Family
ID=68075415
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| PCT/CN2020/087605 Ceased WO2020233368A1 (zh) | 2019-05-22 | 2020-04-28 | 表情识别模型训练方法、装置、设备和存储介质 |
Country Status (2)
| Country | Link |
|---|---|
| CN (1) | CN110309713A (zh) |
| WO (1) | WO2020233368A1 (zh) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112784857A (zh) * | 2021-01-29 | 2021-05-11 | 北京三快在线科技有限公司 | 一种模型训练以及图像处理方法及装置 |
| CN113255517A (zh) * | 2021-05-24 | 2021-08-13 | 中国科学技术大学 | 保护隐私的表情识别模型训练方法及表情识别方法及装置 |
| CN114155370A (zh) * | 2021-10-29 | 2022-03-08 | 北京搜狗科技发展有限公司 | 图像数据处理方法、装置及介质 |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110309713A (zh) * | 2019-05-22 | 2019-10-08 | 深圳壹账通智能科技有限公司 | 表情识别模型训练方法、装置、设备和存储介质 |
| CN111597476B (zh) * | 2020-05-06 | 2023-08-22 | 北京金山云网络技术有限公司 | 图像处理方法及装置 |
| CN116310630A (zh) * | 2023-02-14 | 2023-06-23 | 厦门美图之家科技有限公司 | 调色模型训练方法、处理方法及其装置、设备及存储介质 |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107767416A (zh) * | 2017-09-05 | 2018-03-06 | 华南理工大学 | 一种低分辨率图像中行人朝向的识别方法 |
| CN108492343A (zh) * | 2018-03-28 | 2018-09-04 | 东北大学 | 一种扩充目标识别的训练数据的图像合成方法 |
| US20180286037A1 (en) * | 2017-03-31 | 2018-10-04 | Greg Zaharchuk | Quality of Medical Images Using Multi-Contrast and Deep Learning |
| CN108710831A (zh) * | 2018-04-24 | 2018-10-26 | 华南理工大学 | 一种基于机器视觉的小数据集人脸识别算法 |
| CN110309713A (zh) * | 2019-05-22 | 2019-10-08 | 深圳壹账通智能科技有限公司 | 表情识别模型训练方法、装置、设备和存储介质 |
-
2019
- 2019-05-22 CN CN201910427443.1A patent/CN110309713A/zh active Pending
-
2020
- 2020-04-28 WO PCT/CN2020/087605 patent/WO2020233368A1/zh not_active Ceased
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180286037A1 (en) * | 2017-03-31 | 2018-10-04 | Greg Zaharchuk | Quality of Medical Images Using Multi-Contrast and Deep Learning |
| CN107767416A (zh) * | 2017-09-05 | 2018-03-06 | 华南理工大学 | 一种低分辨率图像中行人朝向的识别方法 |
| CN108492343A (zh) * | 2018-03-28 | 2018-09-04 | 东北大学 | 一种扩充目标识别的训练数据的图像合成方法 |
| CN108710831A (zh) * | 2018-04-24 | 2018-10-26 | 华南理工大学 | 一种基于机器视觉的小数据集人脸识别算法 |
| CN110309713A (zh) * | 2019-05-22 | 2019-10-08 | 深圳壹账通智能科技有限公司 | 表情识别模型训练方法、装置、设备和存储介质 |
Non-Patent Citations (1)
| Title |
|---|
| 有三AI (YOUSAN AI): "一文道尽深度学习中的数据增强方法 (上) (non-official translation: All you Need to Know on Data Enhancement Method in Deep Learning (part 1))", HTTPS://WWW.JIANSHU.COM/P/99450DBDADCF, 28 June 2018 (2018-06-28), DOI: 20200707132046Y * |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112784857A (zh) * | 2021-01-29 | 2021-05-11 | 北京三快在线科技有限公司 | 一种模型训练以及图像处理方法及装置 |
| CN112784857B (zh) * | 2021-01-29 | 2022-11-04 | 北京三快在线科技有限公司 | 一种模型训练以及图像处理方法及装置 |
| CN113255517A (zh) * | 2021-05-24 | 2021-08-13 | 中国科学技术大学 | 保护隐私的表情识别模型训练方法及表情识别方法及装置 |
| CN113255517B (zh) * | 2021-05-24 | 2023-10-24 | 中国科学技术大学 | 保护隐私的表情识别模型训练方法及表情识别方法及装置 |
| CN114155370A (zh) * | 2021-10-29 | 2022-03-08 | 北京搜狗科技发展有限公司 | 图像数据处理方法、装置及介质 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110309713A (zh) | 2019-10-08 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| WO2020233368A1 (zh) | 表情识别模型训练方法、装置、设备和存储介质 | |
| Sun et al. | Improving the stability and efficiency of diffusion models for content consistent super-resolution | |
| CN110969250B (zh) | 一种神经网络训练方法及装置 | |
| CN109408653A (zh) | 基于多特征检索和形变的人体发型生成方法 | |
| CN111259940A (zh) | 一种基于空间注意力地图的目标检测方法 | |
| CN109377445A (zh) | 模型训练方法、替换图像背景的方法、装置和电子系统 | |
| Tian et al. | Kaokore: A pre-modern japanese art facial expression dataset | |
| CN108470320A (zh) | 一种基于cnn的图像风格化方法及系统 | |
| CN110570481A (zh) | 基于风格迁移的书法字库自动修复方法及系统 | |
| CN114495119A (zh) | 一种复杂场景下的实时不规则文本识别方法 | |
| CN117315481A (zh) | 基于光谱-空间自注意力和Transformer网络的高光谱图像分类方法 | |
| CN110727819A (zh) | 一种尺度自适应病理全切片图像数据库检索方法 | |
| CN111783879B (zh) | 基于正交注意力机制的层次化压缩图匹配方法及系统 | |
| CN111368831B (zh) | 一种竖排文字的定位系统及方法 | |
| CN108229576A (zh) | 一种跨倍率病理图像特征学习方法 | |
| CN117557683A (zh) | 一种新型的文本和语义分割图共同驱动的图像生成方法 | |
| CN112733716A (zh) | 基于srocrn网络的低分辨率文本图像识别方法 | |
| CN110929762A (zh) | 一种基于深度学习的肢体语言检测与行为分析方法及系统 | |
| CN114581789A (zh) | 一种高光谱图像分类方法及系统 | |
| CN116458896A (zh) | 一种基于时序特征图与注意力机制的心电图分类方法和装置 | |
| CN117953561B (zh) | 一种基于Transformer和显著图的时空区三流微表情识别方法 | |
| CN115147636A (zh) | 基于胸部x光图像肺部疾病识别和分类方法 | |
| CN106204420A (zh) | 一种笔式图像扫描拼接方法与装置 | |
| CN117274628A (zh) | 结合轮廓波变化和Vision Transformer的图像处理方法 | |
| CN115641323A (zh) | 医学图像自动标注的方法及装置 |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| 121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 20810789 Country of ref document: EP Kind code of ref document: A1 |
|
| NENP | Non-entry into the national phase |
Ref country code: DE |
|
| 122 | Ep: pct application non-entry in european phase |
Ref document number: 20810789 Country of ref document: EP Kind code of ref document: A1 |
|
| 32PN | Ep: public notification in the ep bulletin as address of the adressee cannot be established |
Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 22.03.2022) |
|
| 122 | Ep: pct application non-entry in european phase |
Ref document number: 20810789 Country of ref document: EP Kind code of ref document: A1 |







