News classification
Contact us
- Add: No. 9, North Fourth Ring Road, Haidian District, Beijing. It mainly includes face recognition, living detection, ID card recognition, bank card recognition, business card recognition, license plate recognition, OCR recognition, and intelligent recognition technology.
- Tel: 13146317170 廖经理
- Fax:
- Email: 398017534@qq.com
AI image quality detection method
AI image quality detection method
Image quality detection method
This chapter is simply a traditional one, and there is no way to know how to stop the detection of image content quality.
The
.
1, full, semi-reference method
Some features of the image are compared with the same feature of the original image, such as probability distribution of wavelet transform coefficients, comprehensive multi-scale geometric analysis, sensitivity function of contrast ratio and distinctive feature of gray scale, etc. The corresponding application categories include video transmission. The digital watermark research, application of the secondary channel to stop the video quality monitoring and code rate control.
.
2. Blind image quality (BIQ)
The evaluation method is complete without reference to the image. The quality of the image is estimated based on the characteristics of the distorted image itself. Some methods are oriented to specific types of distortion, such as stopping the evaluation for serious levels of ambiguity, noise, and block effects; some methods stop the distortion first. Reasons for classification, and then stop the quantitative evaluation; and some methods try to evaluate different distortion types of images at the same time. No reference method is the most applicable value, has a very wide range of applications.
.
3, machine learning image quality evaluation
(1) SVM + SVR
A two-step plan is used to calculate the law. The SVM is used to stop the distortion type recognition, and then the SVR regression analysis model is established for the specific distortion type. We call it the SVM + SVR model.
(2) GGD
The blind image quality index (BIQI) of Moorthy and Bovik evaluates the image in two steps. First, the parameters of the wavelet synthesis coefficient fitted by the generalized Gaussian distribution (GGD) model are used as features. The SVM classifies the probability that the current image belongs to each class, then uses SVR to calculate the image quality index value for each degeneracy type, and finally obtains the total quality evaluation index according to the probability weight; In the subsequent image recognition based on distortion identification
Solidity and integrity assessment.
.
4, based on the probability model approach
This type of approach first establishes a statistical probability model between image features and image quality. Most of them use multivariate Gaussian scatter to describe the probability distribution. After evaluating the image, after extracting features, calculate the image quality of the maximum posterior probability according to the probability model, or according to the probability. The matching level of the model (such as the interval between features) measures the image quality.
In the natural image quality evaluator (NIQE) algorithm proposed by Mittal et al. of the University of Texas at Austin, it is not necessary to use the distorted image of the human eye to stop the exercise, after calculating some of the MSCN normalized images. Based on partial activity, local image blocks are selected as exercise data, model parameters are fitted by generalized Gaussian model as features, multivariate Gaussian models are used to describe these features, and the parameters of the image feature model to be evaluated and the model parameters set in advance are used in the evaluation process. The interval to confirm the image quality
Abdalmajeed and Jiao, after normalizing the image stop portion MSCN, extract the natural image statistical features based on Weber scattering and plot its probability distribution with a multivariate Gaussian spread, and calculate the interval between the statistical model of the image to be evaluated and the undistorted image statistically. As a measure of image quality evaluation, modeling based on probability is a statistical method based on a large number of samples. The choice of probability mathematical model and the size of sample size are the keys to performance. Existing methods are mostly based on multivariate Gaussian model stopping probability modeling. It is mainly for easy modeling. It is thought that the characteristic dimension for characterizing image quality is high, and the complex model will require more data. Such methods can only achieve better results when the data volume is large.
.
5, neural network approach
This kind of method first extracts certain image transform domain or spatial features, then exercises a neural network regression parsing model based on known quality data, and predicts image quality by image features.
Kang et al. used convolutional neural networks (CNN) to integrate feature extraction and regression analysis into the same network. The network consists of 5 layers. The image is normalized by some MSCNs and then input into the network with 32 £32 sub-blocks. The first layer The convolutional layer extracts features by 50 filters, the second layer stops the maximum and minimum selections, the next two layers are 800-node fully meshed networks, and the last layer outputs the image quality for a single node.
Hou et al. also adopted a deep learning algorithm with a 5-layer network structure to stop the image quality evaluation, integrated features extraction, classification, and posterior probability calculations into one function. The detailed features of the 3-level wavelet transform were used as inputs, and the training process used restricted waves first. Restricted Boltzmann machine (RBM) stops inter-layer learning, and then uses the reverse transfer algorithm to stop fine-tuning. The experimental results of these two algorithms are significantly better than other non-reference algorithms, and in some cases are better than others. Good VIF in full-reference algorithm
The
This chapter is simply a traditional one, and there is no way to know how to stop the detection of image content quality.
The
.
1, full, semi-reference method
Some features of the image are compared with the same feature of the original image, such as probability distribution of wavelet transform coefficients, comprehensive multi-scale geometric analysis, sensitivity function of contrast ratio and distinctive feature of gray scale, etc. The corresponding application categories include video transmission. The digital watermark research, application of the secondary channel to stop the video quality monitoring and code rate control.
.
2. Blind image quality (BIQ)
The evaluation method is complete without reference to the image. The quality of the image is estimated based on the characteristics of the distorted image itself. Some methods are oriented to specific types of distortion, such as stopping the evaluation for serious levels of ambiguity, noise, and block effects; some methods stop the distortion first. Reasons for classification, and then stop the quantitative evaluation; and some methods try to evaluate different distortion types of images at the same time. No reference method is the most applicable value, has a very wide range of applications.
.
3, machine learning image quality evaluation
(1) SVM + SVR
A two-step plan is used to calculate the law. The SVM is used to stop the distortion type recognition, and then the SVR regression analysis model is established for the specific distortion type. We call it the SVM + SVR model.
(2) GGD
The blind image quality index (BIQI) of Moorthy and Bovik evaluates the image in two steps. First, the parameters of the wavelet synthesis coefficient fitted by the generalized Gaussian distribution (GGD) model are used as features. The SVM classifies the probability that the current image belongs to each class, then uses SVR to calculate the image quality index value for each degeneracy type, and finally obtains the total quality evaluation index according to the probability weight; In the subsequent image recognition based on distortion identification
Solidity and integrity assessment.
.
4, based on the probability model approach
This type of approach first establishes a statistical probability model between image features and image quality. Most of them use multivariate Gaussian scatter to describe the probability distribution. After evaluating the image, after extracting features, calculate the image quality of the maximum posterior probability according to the probability model, or according to the probability. The matching level of the model (such as the interval between features) measures the image quality.
In the natural image quality evaluator (NIQE) algorithm proposed by Mittal et al. of the University of Texas at Austin, it is not necessary to use the distorted image of the human eye to stop the exercise, after calculating some of the MSCN normalized images. Based on partial activity, local image blocks are selected as exercise data, model parameters are fitted by generalized Gaussian model as features, multivariate Gaussian models are used to describe these features, and the parameters of the image feature model to be evaluated and the model parameters set in advance are used in the evaluation process. The interval to confirm the image quality
Abdalmajeed and Jiao, after normalizing the image stop portion MSCN, extract the natural image statistical features based on Weber scattering and plot its probability distribution with a multivariate Gaussian spread, and calculate the interval between the statistical model of the image to be evaluated and the undistorted image statistically. As a measure of image quality evaluation, modeling based on probability is a statistical method based on a large number of samples. The choice of probability mathematical model and the size of sample size are the keys to performance. Existing methods are mostly based on multivariate Gaussian model stopping probability modeling. It is mainly for easy modeling. It is thought that the characteristic dimension for characterizing image quality is high, and the complex model will require more data. Such methods can only achieve better results when the data volume is large.
.
5, neural network approach
This kind of method first extracts certain image transform domain or spatial features, then exercises a neural network regression parsing model based on known quality data, and predicts image quality by image features.
Kang et al. used convolutional neural networks (CNN) to integrate feature extraction and regression analysis into the same network. The network consists of 5 layers. The image is normalized by some MSCNs and then input into the network with 32 £32 sub-blocks. The first layer The convolutional layer extracts features by 50 filters, the second layer stops the maximum and minimum selections, the next two layers are 800-node fully meshed networks, and the last layer outputs the image quality for a single node.
Hou et al. also adopted a deep learning algorithm with a 5-layer network structure to stop the image quality evaluation, integrated features extraction, classification, and posterior probability calculations into one function. The detailed features of the 3-level wavelet transform were used as inputs, and the training process used restricted waves first. Restricted Boltzmann machine (RBM) stops inter-layer learning, and then uses the reverse transfer algorithm to stop fine-tuning. The experimental results of these two algorithms are significantly better than other non-reference algorithms, and in some cases are better than others. Good VIF in full-reference algorithm
The