| Image distortion widely exists in image acquisition,transmission and processing.Evaluating the quality of an image has important guiding significance for optimizing the image quality.In recent years,it has gradually become a hot research topic in image processing and artificial intelligence.Considering that human eyes are the terminal in most application scenes,this paper aims to improve the prediction accuracy and efficiency of image quality assessment(IQA)methods by studying the perception process of human visual system(HVS).From theoretical perspective,aiming at the problems of existing methods in contrast quality perception,image representation,perception feature extraction,image perception process modeling etc.,a IQA method based on contrast distortion characteristics,an unsupervised IQA method based on quaternion representation,an unsupervised IQA method based on multi-scale structure features and an unsupervised IQA method based on top-down perception mechanism are established.From application perspective,aiming at the problems that existing working condition recognition methods has low prediction accuracy and poor robustness when the flame image has low quality,a high robustness and high-precision sintering condition recognition method based on the quality of flame video image is proposed.This paper has following innovations:(1)Based on the inherent law of contrast-distorted image,a high-performance IQA method for contrast distorted image is proposed.First,based on the fact that contrast distortion changes the frequency band of the image,the singular values of the image are selected as the detection feature of contrast distortion.Second,by analyzing the influence of different distortion parameters on contrast image quality,it is found that the contrast distortion parameters and image quality scores are asymmetrically distributed.Last,based on this characteristic,an adaptive nonlinear fitting model is designed to convert features into quality score.The experimental results on the contrast image database show that compared with the existing methods,the proposed method not only obtains higher prediction accuracy,but also has great advantages in computational complexity.(2)Aiming at the problem that the existing unsupervised methods adopt the single channel evaluation strategy and do not consider the information coupling between channels,an unsupervised IQA method based on quaternion representation is proposed.The traditional color image processing method is to convert the color image into gray one or process the three channels of the color image respectively before fusion.Such methods do not take the correlation between color channelsinto into account.Based on this concern,this paper represents the color image with a quaternion,and designs quaternion MSCN operator,gradient operator and Gabor operator to extract naturalness,structure and texture features,so as to realize multi-channel fusion perception of image quality.The experimental results on the color image database show that this method not only reduces the computational complexity,but also effectively extracts the correlation between the color channels and improves the prediction accuracy of the image quality evaluation model.(3)Referring to the multi-scale receptive field structure of HVS,and aiming at the current situation that the existing unsupervised methods only extract the pixel level scale features from the image,this paper proposes an unsupervised IQA method by extracting multi-scale structure features.In the process of HVS perception,the large and small cells of the lateral geniculate body have different receptive fields to sense the large and small-scale information in the visual field.However,the existing unsupervised IQA methods only extract local perceptual features at the pixel-level of the image.This kind of features can only simulate the feature extraction method of human small cells.The perceptual process of large cells is ignored.To this end,this paper proposes a large-scale texture feature extraction method from the image patch.Considering that the structure and texture intensity of an image are directly proportional to its high-frequency energy,the ratio of the high-frequency energy is proposed as the texture intensity feature of the patch.The small-scale features and large-scale features are combined to simulate the information processing mode of lateral geniculate body.The experimental results on the image database show that compared with the best unsupervised image quality evaluation method,this method not only obtains higher accuracy,but also has lower computational complexity.(4)Referring to the top-down image perception process,an unsupervised IQA method based on visual perception mechanism is proposed.HVS will first recognize the content in the image,and then focus on the local area for local quality perception,resulting in the global perception and local perception process.Existing unsupervised methods only perceive local quality from the image,ignoring the global quality perception.For this reason,by introducing an image object detection confidence-based global quality perception model,this paper proposes a global and local perception combined framework for unsupervised IQA.At the feature extraction level,in view of the complementary characteristics of various types of features in image representation,it is proposed to combine fitting and histogram features that are extracted based on prior knowledge with hidden features based on deep neural network learning to achieve an all-round perception of distortion.Experiments on the image database show that the proposed method has higher prediction accuracy compared with the traditional full reference method PSNR.(5)The above-mentioned IQA methods are applied to the condition recognition of high-temperature sintering process based on coal-burning flame images.By studying the relationship between the quality sequence of flame video and working conditions,this paper proposes a robust sintering condition recognition method based on flame image quality evaluation.First,aiming at the problem that the reference image,labels and high-quality flame image are unavailable in flame image quality evaluation,a guided filtering method is adopted to obtain high-definition flame image,which can be used as reference information for blurred flame images to achieve flame image quality evaluation.Second,the quality sequence of flame video is used as the input of the flame condition recognition model,and mc ODM is used to recognize the flame condition.Experiments on real flame video show the effectiveness and stability of the proposed method. |
