CN106651830A - Image quality test method based on parallel convolutional neural network - Google Patents

本发明公开了一种基于并行卷积神经网络的图像质量测试方法，包括以下步骤：(1)采用并行卷积神经网络建立图像质量测试模型；所述图像质量测试模型包括五个卷积层和三个全连接层；(2)输入数据预处理与数据库平衡化处理：(3)模型的预训练：采用预训练数据集，对图像质量测试模型进行预训练学习，得到网络权值；(4)并行模型训练：初始化图像质量测试模型，基于预训练初始化后的图像质量测试模型，进行并行模型训练，得到已训练的图像质量评估模型；(5)对目标图像使用已训练的质量评估模型进行测试。本发明得到的测试结果符合人类审美标准，判断过程无需人工参与，实现了机器全自动图像质量评价。

The invention discloses an image quality testing method based on a parallel convolutional neural network, comprising the following steps: (1) establishing an image quality testing model by using a parallel convolutional neural network; the image quality testing model includes five convolutional layers and Three fully connected layers; (2) input data preprocessing and database balancing processing: (3) model pre-training: using pre-training data sets, pre-training and learning the image quality test model to obtain network weights; (4) ) Parallel model training: initialize the image quality test model, perform parallel model training based on the image quality test model after pre-training initialization, and obtain the trained image quality assessment model; (5) use the trained quality assessment model for the target image test. The test results obtained by the invention conform to human aesthetic standards, the judging process does not require manual participation, and the automatic machine image quality evaluation is realized.

一种基于并行卷积神经网络的图像质量测试方法An Image Quality Testing Method Based on Parallel Convolutional Neural Network

技术领域technical field

本发明涉及图像信号处理领域，特别涉及一种基于并行卷积神经网络的图像质量测试方法。The invention relates to the field of image signal processing, in particular to an image quality testing method based on a parallel convolutional neural network.

背景技术Background technique

人类的美感感受和判断虽然受到文化背景、个人经历、时代背景等的影响，但是总体上具有很大的共性。无数绘画、摄影和艺术作品作为人类共同的审美财富普遍受到人们的欣赏和喜爱。美学质量评估就是希望通过计算机，模拟人类高层感知来判断图像的美感，实现对图像进行高质量或低质量分类，或者对图像的质量程度给出评分。Although human beings' aesthetic feelings and judgments are influenced by cultural background, personal experience, time background, etc., they generally have a lot of commonality. Numerous paintings, photographs and works of art are generally appreciated and loved by people as the common aesthetic wealth of mankind. Aesthetic quality assessment is to use computers to simulate the high-level perception of human beings to judge the beauty of images, to classify images as high-quality or low-quality, or to score the quality of images.

传统的图像质量评估方法大多采用手选识别特征的方式，图像特征的有效提取对分类结果具有至关重要的作用。例如尝试借鉴摄影，艺术，绘画等领域的规则、人类审美经验、视觉注意机制，从图像中提取各种各样的图像特征，例如边缘特征，颜色直方图特征，三分法则特征等等。还有一些使用局部特征的方法，例如SIFT(Scale-invariant featuretransform)算法，词袋(Bag Of Words，Bow)算法，FisherVector(FV)算法，或者它们的改进算法等。这些方法都取得了较好的应用价值。Most of the traditional image quality assessment methods use hand-selected recognition features, and the effective extraction of image features plays a vital role in the classification results. For example, try to learn from the rules of photography, art, painting and other fields, human aesthetic experience, and visual attention mechanism to extract various image features from images, such as edge features, color histogram features, rule of thirds features, and so on. There are also methods using local features, such as SIFT (Scale-invariant featuretransform) algorithm, Bag Of Words (Bow) algorithm, FisherVector (FV) algorithm, or their improved algorithms. These methods have achieved good application value.

深度学习在解决传统计算机视觉问题上有突破性进展，尤其是卷积神经网络(Convolutional Neural Network，CNN)的应用。通过直接利用大量的数据训练多层CNN，不需要先验知识和经验，人们发现网络对于学习到的特征具有较好的鲁棒性，不仅省去了复杂繁琐的手动特征提取过程，更能从样本中发现更为重要并难以理解的高层特征。利用深度学习进行图像质量评价研究中，宾夕法尼亚州立大学的学者Wang等人设计了一个双通道的卷积神经网络用于图像质量分类人物。中国科学技术大学的田教授等人利用深度学习网络来提取图像特征，然后使用支持向量机(Support Vector Machine，SVM)进行图像质量分类。这些是深度学习方法在图像质量评估方面的初步尝试，取得了一定的效果。Deep learning has made breakthroughs in solving traditional computer vision problems, especially the application of Convolutional Neural Network (CNN). By directly using a large amount of data to train multi-layer CNN without prior knowledge and experience, people found that the network has better robustness to the learned features, which not only saves the complicated and cumbersome manual feature extraction process, but also can learn from More important and difficult to understand high-level features found in the sample. In the study of image quality evaluation using deep learning, Wang et al., a scholar at Pennsylvania State University, designed a two-channel convolutional neural network for image quality classification of people. Professor Tian of the University of Science and Technology of China and others used a deep learning network to extract image features, and then used a Support Vector Machine (Support Vector Machine, SVM) to classify image quality. These are the initial attempts of deep learning methods in image quality assessment and have achieved certain results.

然而不同场景类别的图像差异大，这导致不同图像特征对于不同场景类别图像的适应性较差。另外，图像的一些复杂的构图规则和质量评估规律在工程上难以被建模和量化，这成为图像特征提取上的瓶颈。However, images of different scene categories are very different, which leads to poor adaptability of different image features to images of different scene categories. In addition, some complex composition rules and quality evaluation rules of images are difficult to be modeled and quantified in engineering, which has become a bottleneck in image feature extraction.

因此需要一种新的测试模型来克服现有技术中存在的问题。Therefore, a new test model is needed to overcome the problems in the prior art.

发明内容Contents of the invention

为了克服现有技术的上述缺点与不足，本发明的目的在于提供一种基于并行卷积神经网络的图像质量测试方法，克服了传统方法需要手工设计多种图像特征的缺点，深入分析和挖掘图像质量特征，泛化能力强，分类准确率高。In order to overcome the above-mentioned shortcomings and deficiencies of the prior art, the object of the present invention is to provide an image quality testing method based on a parallel convolutional neural network, which overcomes the shortcomings of traditional methods that require manual design of various image features, and deeply analyzes and mines images. Quality features, strong generalization ability, and high classification accuracy.

本发明的目的通过以下技术方案实现：The object of the present invention is achieved through the following technical solutions:

一种基于并行卷积神经网络的图像质量测试方法，包括以下步骤：A method for testing image quality based on a parallel convolutional neural network, comprising the following steps:

(1)采用并行卷积神经网络建立图像质量测试模型；所述图像质量测试模型包括第一卷积层、第二卷积层、第三卷积层、第四卷积层、第五卷积层、第一全连接层、第二全连接层和第三全连接层；(1) adopt parallel convolutional neural network to set up image quality testing model; Described image quality testing model comprises first convolutional layer, second convolutional layer, the 3rd convolutional layer, the 4th convolutional layer, the 5th convolutional layer layer, the first fully connected layer, the second fully connected layer and the third fully connected layer;

所述第五卷积层为包含n个分支的并行结构网络；1≤n≤10；The fifth convolutional layer is a parallel structure network including n branches; 1≤n≤10;

(2)输入数据预处理与数据库平衡化处理：对预训练数据集的每个样本进行裁剪和归一化，并对预训练数据集的样本数量进行平衡化处理；(2) Input data preprocessing and database balance processing: each sample of the pre-training data set is cut and normalized, and the number of samples of the pre-training data set is balanced;

(3)模型的预训练：采用预训练数据集，对图像质量测试模型进行预训练学习，得到网络权值；(3) Pre-training of the model: using the pre-training data set, the image quality test model is pre-trained and learned to obtain the network weight;

所述预训练学习，具体为：The pre-training study is specifically:

用预训练数据集中每一种类别图像各自训练一个深度CNN网络，并且进行权值学习和提取；Use each category of images in the pre-training data set to train a deep CNN network, and perform weight learning and extraction;

所述权值学习和提取，具体包括以下步骤：The weight learning and extraction specifically includes the following steps:

(3-1)深度CNN网络权值初始化；(3-1) Deep CNN network weight initialization;

(3-2)对深度CNN网络进行迭代训练；(3-2) Iteratively train the deep CNN network;

(3-3)提取每一个深度CNN网络第五卷积层学习得到的卷积核权值；(3-3) Extract the convolution kernel weights learned by the fifth convolutional layer of each deep CNN network;

(4)并行模型训练：初始化图像质量测试模型，基于预训练初始化后的图像质量测试模型，进行并行模型训练，得到已训练的图像质量评估模型；(4) Parallel model training: initialize the image quality test model, perform parallel model training based on the image quality test model after pre-training initialization, and obtain the trained image quality evaluation model;

(5)对目标图像使用已训练的质量评估模型进行测试。(5) Test the target image using the trained quality assessment model.

步骤(2)所述输入数据预处理，具体为：将所有的样本的长宽统一归一化为256*256，在模型输入接口的匹配中，每一次读取输入图像数据时，被随机裁剪到规格为227*227的大小。The input data preprocessing described in step (2) is specifically: the length and width of all samples are uniformly normalized to 256*256, and in the matching of the model input interface, each time the input image data is read, it is randomly cropped to the size of 227*227.

步骤(2)所述数据库平衡化处理，具体为：对预训练数据集中的每个样本进行旋转处理，并左右镜像一次，产生新的样本。The database balancing process described in step (2) specifically includes: performing rotation processing on each sample in the pre-training data set, and mirroring it left and right once to generate a new sample.

步骤(3)所述预训练学习中，第一～第四卷积层初始设为AlexNet模型前四层网络权值，采用随即梯度下降方式进行训练，学习率设置为初始值0.0001，第五卷积层提取层和全连接层则设置初始学习率为0.001。In the pre-training learning described in step (3), the first to fourth convolutional layers are initially set to the network weights of the first four layers of the AlexNet model, and the training is carried out using a random gradient descent method, and the learning rate is set to an initial value of 0.0001. The initial learning rate of the multilayer extraction layer and the fully connected layer is set to 0.001.

步骤(4)所述初始化图像质量测试模型，基于预训练初始化后的图像质量测试模型，进行并行模型训练：Step (4) described initialization image quality test model, based on the image quality test model after pre-training initialization, carry out parallel model training:

(4-1)模型初始化；(4-1) Model initialization;

(4-2)设置训练参数；(4-2) Set training parameters;

(4-3)加载训练数据，所述训练数据包括训练集和验证集；(4-3) load training data, described training data comprises training set and verification set;

(4-4)采用随即梯度下降算法对初始化后的图像质量测试模型进行迭代训练，在训练集上，每迭代1000次保存一次模型参数，经过不断迭代，取得网络最优解，取在验证集上误差最小的模型作为已训练的图像质量评估模型。(4-4) Use the random gradient descent algorithm to iteratively train the initialized image quality test model. On the training set, save the model parameters every 1000 iterations. After continuous iterations, obtain the optimal solution of the network and take it in the verification set. The model with the smallest error above is used as the trained image quality assessment model.

步骤(4-1)所述模型初始化，具体为：The model initialization described in step (4-1) is specifically:

引用AlexNet模型的权值来初始化图像质量测试模型的第一～第四卷积层，第五层的并行结构由步骤(3)中预训练阶段得到的权值进行初始化，全连接层的则权值采用随机初始化方式。Citing the weights of the AlexNet model to initialize the first to fourth convolutional layers of the image quality test model, the parallel structure of the fifth layer is initialized by the weights obtained in the pre-training stage in step (3), and the weights of the fully connected layer are Values are initialized randomly.

步骤(4-2)所述设置训练参数，具体为：The training parameters described in step (4-2) are set, specifically:

第一～第五卷积层的初始学习率设置为0.0001；全连接层参数的初始学习率为0.001；训练过程设为每8次遍历样本集后，学习率降低40％。The initial learning rate of the first to fifth convolutional layers is set to 0.0001; the initial learning rate of the parameters of the fully connected layer is 0.001; the training process is set to reduce the learning rate by 40% after traversing the sample set every 8 times.

所述深度网络模型结构，具体如下：The structure of the deep network model is as follows:

第一卷积层有96个卷积核，大小为11*11*3；第二卷积层有256个卷积核，大小为5*5*48；第三卷积层有384个核，大小为3*3*256；第四卷积层有384个核，大小为3*3*192；第5卷积层有64*n个核，大小为3*3*64；第一和第二全连接层有512和神经元，第三全连接层有2个神经元；The first convolution layer has 96 convolution kernels with a size of 11*11*3; the second convolution layer has 256 convolution kernels with a size of 5*5*48; the third convolution layer has 384 kernels, The size is 3*3*256; the fourth convolutional layer has 384 cores, the size is 3*3*192; the fifth convolutional layer has 64*n cores, the size is 3*3*64; the first and the first The second fully connected layer has 512 and neurons, and the third fully connected layer has 2 neurons;

第一层卷积层依次经第一池化层、第一正则化层与第二卷积层连接；第二卷积层经第二池化层、第二正则化层与第三卷积层连接；第一池化层、第二池化层参数与AlexNet模型参数相同；第三卷积层直接与第四卷积层连接；第四卷积层直接与第五卷积层连接；第五卷积层经第五池化层与第一全连接层连接，第五池化层采用均值池化方法，池化单元大小z*z取2*2，池化步长s取2；第一全连接层依次连接第二全连接层和第三全连接层。The first convolutional layer is sequentially connected through the first pooling layer, the first regularization layer and the second convolutional layer; the second convolutional layer is connected through the second pooling layer, the second regularization layer and the third convolutional layer Connection; the parameters of the first pooling layer and the second pooling layer are the same as those of the AlexNet model; the third convolutional layer is directly connected to the fourth convolutional layer; the fourth convolutional layer is directly connected to the fifth convolutional layer; the fifth The convolutional layer is connected to the first fully connected layer through the fifth pooling layer, the fifth pooling layer adopts the mean pooling method, the pooling unit size z*z is 2*2, and the pooling step s is 2; the first The fully connected layer is sequentially connected to the second fully connected layer and the third fully connected layer.

与现有技术相比，本发明具有以下优点和有益效果：Compared with the prior art, the present invention has the following advantages and beneficial effects:

(1)本发明采用并行深度卷积神经网络模型，并行多个独立分支，有效结合了传统的特征融合方法，具有很好的可扩展性，同时提高了模型的特征表达能力。(1) The present invention adopts a parallel deep convolutional neural network model, parallels multiple independent branches, effectively combines traditional feature fusion methods, has good scalability, and improves the feature expression ability of the model at the same time.

(2)本发明提出的预训练过程中具有先进性，具体表现在：选取噪声小、干净的数据集作为模型预训练数据；第五卷积层多个分支可以全面的学习到图像质量信息；借用现今模型的优点进一步优化提升模型性能。(2) The pre-training process proposed by the present invention is advanced, which is embodied in: selecting a low-noise, clean data set as model pre-training data; multiple branches of the fifth convolutional layer can comprehensively learn image quality information; Borrow the advantages of the current model to further optimize and improve the performance of the model.

(3)在全连接层和卷积层之间使用均值池化层，降低噪声对特征数据的影响，增强分类效果。(3) Use the mean pooling layer between the fully connected layer and the convolutional layer to reduce the influence of noise on the feature data and enhance the classification effect.

(4)提出一种利用多种数据集完成图像质量分类系统的训练和测试方法，有利于在大数据量的图像库中进行快速质量分类，方法简单有效，可靠性高。(4) A training and testing method for image quality classification system using multiple data sets is proposed, which is conducive to rapid quality classification in image databases with large amounts of data. The method is simple, effective, and highly reliable.

附图说明Description of drawings

图1为本发明的实施例的基于并行卷积神经网络的图像质量测试方法的训练和工作流程图。FIG. 1 is a training and working flow chart of an image quality testing method based on a parallel convolutional neural network according to an embodiment of the present invention.

图2为本发明的图像质量测试模型结构图。Fig. 2 is a structural diagram of the image quality testing model of the present invention.

图3为本发明的在预训练阶段用于权值学习的卷积神经网络模型结构图。FIG. 3 is a structural diagram of a convolutional neural network model used for weight learning in the pre-training stage of the present invention.

具体实施方式detailed description

下面结合实施例，对本发明作进一步地详细说明，但本发明的实施方式不限于此。The present invention will be described in further detail below in conjunction with the examples, but the embodiments of the present invention are not limited thereto.

实施例Example

如图1所示，本实施例的基于并行卷积神经网络的图像质量测试方法，包括以下步骤：As shown in Figure 1, the image quality testing method based on parallel convolutional neural network of the present embodiment comprises the following steps:

(1)采用并行卷积神经网络建立图像质量测试模型；所述图像质量测试模型包括第一卷积层、第二卷积层、第三卷积层、第四卷积层、第五卷积层、第一全连接层、第二全连接层和第三全连接层；所述第五卷积层为包含n个分支的并行结构网络；1≤n≤10。(1) adopt parallel convolutional neural network to set up image quality testing model; Described image quality testing model comprises first convolutional layer, second convolutional layer, the 3rd convolutional layer, the 4th convolutional layer, the 5th convolutional layer layer, the first fully connected layer, the second fully connected layer and the third fully connected layer; the fifth convolutional layer is a parallel structure network including n branches; 1≤n≤10.

如图2所示，本实施例的图像质量测试模型一个包含5层卷积层和3个全连接层的8层的深度卷积神经网络，此模型前四层卷积层借用了Alexnet[A.Krizhevsky,I.Sutskever,G.E.Hinton,ImageNet classification with deep convolution neuralnetworks,in:Proceedings of the Annual Conference on Neural InformationProcessing System(NIPS),2012,pp.1097-1105.]的前四层网络结构与参数。第五层定义为场景卷积层，本实施例由7组卷积核并行构成，用于学习不同场景类别下的图像特征。每个分支分别连接第四层与第五层中的一组卷积网络。As shown in Figure 2, the image quality test model of the present embodiment is an 8-layer deep convolutional neural network comprising 5 layers of convolutional layers and 3 fully connected layers, and the first four layers of convolutional layers of this model borrow Alexnet [A .Krizhevsky, I.Sutskever, G.E.Hinton, ImageNet classification with deep convolution neural networks, in: Proceedings of the Annual Conference on Neural Information Processing System (NIPS), 2012, pp.1097-1105.] The network structure and parameters of the first four layers. The fifth layer is defined as the scene convolution layer. In this embodiment, 7 groups of convolution kernels are formed in parallel to learn image features under different scene categories. Each branch connects a set of convolutional networks in the fourth and fifth layers respectively.

第一卷积层有96个卷积核，大小为11*11*3；第二卷积层有256个卷积核，大小为5*5*48；第三卷积层有384个核，大小为3*3*256；第四卷积层有384个核，大小为3*3*192；第5卷积层有64*n个核，大小为3*3*64；第一和第二全连接层有512和神经元，第三全连接层有2个神经元；The first convolution layer has 96 convolution kernels with a size of 11*11*3; the second convolution layer has 256 convolution kernels with a size of 5*5*48; the third convolution layer has 384 kernels, The size is 3*3*256; the fourth convolutional layer has 384 cores, the size is 3*3*192; the fifth convolutional layer has 64*n cores, the size is 3*3*64; the first and the first The second fully connected layer has 512 and neurons, and the third fully connected layer has 2 neurons;

第一层卷积层依次经第一池化层、第一正则化层与第二卷积层连接；第二卷积层经第二池化层、第二正则化层与第三卷积层连接；第一池化层、第二池化层参数与AlexNet模型参数相同；第三卷积层直接与第四卷积层连接；第四卷积层直接与第五卷积层连接；第五卷积层经第五池化层与第一全连接层连接，第五池化层采用均值池化方法，池化单元大小z*z取2*2，池化步长s取2；第一全连接层依次连接第二和第三全连接层；The first convolutional layer is sequentially connected through the first pooling layer, the first regularization layer and the second convolutional layer; the second convolutional layer is connected through the second pooling layer, the second regularization layer and the third convolutional layer Connection; the parameters of the first pooling layer and the second pooling layer are the same as those of the AlexNet model; the third convolutional layer is directly connected to the fourth convolutional layer; the fourth convolutional layer is directly connected to the fifth convolutional layer; the fifth The convolutional layer is connected to the first fully connected layer through the fifth pooling layer, the fifth pooling layer adopts the mean pooling method, the pooling unit size z*z is 2*2, and the pooling step s is 2; the first The fully connected layer is sequentially connected to the second and third fully connected layers;

(2)输入数据预处理与数据库平衡化处理：对预训练数据集的每个样本进行裁剪和归一化，并对预训练数据集的样本数量进行平衡化处理；(2) Input data preprocessing and database balance processing: each sample of the pre-training data set is cut and normalized, and the number of samples of the pre-training data set is balanced;

步骤(2)所述输入数据预处理，具体为：The input data preprocessing described in step (2) is specifically:

将所有的样本的长宽统一归一化为256*256，在模型输入接口的匹配中，每一次读取输入图像数据时，被随机裁剪到规格为227*227的大小。通过这样的方式，确保不会丢失图像的全局信息。The length and width of all samples are uniformly normalized to 256*256. In the matching of the model input interface, each time the input image data is read, it is randomly cropped to a size of 227*227. In this way, it is ensured that the global information of the image is not lost.

不平衡的训练数据集会对分类结果产生不良影响，弱化学习得到特征的表达能力。预训练阶段用到的CUHKPQ数据库包含有17690张图片，图片集一共有7个类别，分别是"animal","plant","static","architecture","landscape","human"和"night"。每一种类别图片都标有相同的两个标签，高质量和低质量。这个数据集噪声小，被选用来做预训练阶段的训练数据。此外，由于该数据集不平衡，CUHKPQ数据集的高质量图片与低质量图片之比大约是1比3，在把数据分为训练集和测试集之后，本发明所提方法对训练集做了平衡化处理，以确保预训练得到模型的有效性。具体做法如下：An unbalanced training data set will adversely affect the classification results and weaken the expressive ability of the learned features. The CUHKPQ database used in the pre-training phase contains 17,690 pictures. There are 7 categories in the picture set, namely "animal", "plant", "static", "architecture", "landscape", "human" and "night ". Each category of images is marked with the same two labels, high quality and low quality. This data set has low noise and is selected as the training data for the pre-training phase. In addition, due to the unbalanced data set, the ratio of high-quality pictures to low-quality pictures in the CUHKPQ data set is about 1 to 3. After the data is divided into training set and test set, the method proposed in the present invention makes Balance processing to ensure the effectiveness of the pre-trained model. The specific method is as follows:

对训练集中的每张高质量图片进行旋转270°处理，并左右镜像一次，产生两张额外新的样本。使高质量图像的数量达到和低质量图片数量大致相等。Each high-quality image in the training set is rotated 270° and mirrored left and right to generate two additional new samples. Make the number of high-quality images approximately equal to the number of low-quality images.

(3)模型的预训练：采用预训练数据集，对图像质量测试模型进行预训练学习，得到网络权值；(3) Pre-training of the model: using the pre-training data set, the image quality test model is pre-trained and learned to obtain the network weight;

所述预训练学习，具体为：The pre-training study is specifically:

用预训练数据集中每一种类别图像各自训练一个深度CNN网络，并且进行权值学习和提取；预训练学习中，第一～第四卷积层初始设为AlexNet模型训练后的参数后，采用随即梯度下降方式进行训练，学习率设置为初始值0.0001，第五卷积层提取层和全连接层则设置初始学习率为0.001。Use each type of image in the pre-training data set to train a deep CNN network, and perform weight learning and extraction; in the pre-training study, the first to fourth convolutional layers are initially set to the parameters after AlexNet model training, and then use Then the gradient descent method is used for training, the learning rate is set to an initial value of 0.0001, and the initial learning rate of the fifth convolutional layer extraction layer and the fully connected layer is set to 0.001.

所述权值学习和提取，具体包括以下步骤：The weight learning and extraction specifically includes the following steps:

(3-1)深度CNN网络权值初始化；(3-1) Deep CNN network weight initialization;

(3-2)对深度CNN网络进行迭代训练；(3-2) Iteratively train the deep CNN network;

(3-3)提取每一个深度CNN网络第五卷积层学习得到的卷积核权值；(3-3) Extract the convolution kernel weights learned by the fifth convolutional layer of each deep CNN network;

图3所示为用于权值学习的卷积神经网络模型结构图。对于每种场景类型图像，分别进行场景图像特征的学习。训练时是单通道深度学习网络结构，前四层卷积层与图3的卷积层一样，第五层为图3第五层的一个卷积组，全连接层的神经元个数为512，网络的最后一层是2个神经元连接着Softmax函数作为输出。它表示输入图像是属于高质量或低质量类别。Figure 3 shows the structure diagram of the convolutional neural network model for weight learning. For each scene type image, the scene image features are learned separately. The training is a single-channel deep learning network structure. The first four convolutional layers are the same as those in Figure 3. The fifth layer is a convolutional group of the fifth layer in Figure 3. The number of neurons in the fully connected layer is 512. , the last layer of the network is 2 neurons connected with Softmax function as output. It indicates whether the input image belongs to the high-quality or low-quality category.

在本实施例中，将第5层学习到的权值表征为场景图像特征。这样，一一用图3所示网络对7种类别的图像进行训练和学习，取出第五层的7组学习到的卷积核权值，用这些权值初始化并行网络的第五层网络，完成了模型预训练过程。In this embodiment, the weight values learned in the fifth layer are represented as scene image features. In this way, one by one, use the network shown in Figure 3 to train and learn 7 categories of images, take out 7 sets of convolution kernel weights learned in the fifth layer, and use these weights to initialize the fifth layer of the parallel network. The model pre-training process is completed.

在深度CNN网络的学习阶段中，采用基本的Softmax计算损失函数，在而分类任务的情况下，变换为简单的逻辑回归函数，图像输入为x，标签为y。损失函数的计算如下列公式所示：In the learning phase of the deep CNN network, the basic Softmax is used to calculate the loss function, and in the case of the classification task, it is transformed into a simple logistic regression function with the image input as x and the label as y. The calculation of the loss function is shown in the following formula:

其中，m表示为图片数量，预测函数h_θ(x_i)的表达是：Among them, m represents the number of pictures, and the expression of the prediction function h _θ ( _xi ) is:

其中，x_i为第i张输入图像，y_i为第i输入图像所对应的标签数据。Among them, x _i is the i-th input image, and y _i is the label data corresponding to the i-th input image.

(4)并行模型训练：初始化图像质量测试模型，基于预训练初始化后的图像质量测试模型，进行并行模型训练，得到已训练的图像质量评估模型，具体步骤如下：(4) Parallel model training: Initialize the image quality test model, perform parallel model training based on the image quality test model after pre-training initialization, and obtain the trained image quality evaluation model, the specific steps are as follows:

(4-1)模型初始化：引用AlexNet模型的权值来初始化图像质量测试模型的第一～第四卷积层，第五层的并行结构由步骤3中预训练阶段得到的权值进行初始化，全连接层的则权值采用随机初始化方式；(4-1) Model initialization: refer to the weights of the AlexNet model to initialize the first to fourth convolutional layers of the image quality test model, and the parallel structure of the fifth layer is initialized by the weights obtained in the pre-training stage in step 3. The weights of the fully connected layer are initialized randomly;

(4-2)设置训练参数：第一～第五卷积层的初始学习率设置为0.0001；全连接层参数的初始学习率为0.001；训练过程设为每8次遍历样本集后，学习率降低40％；(4-2) Set training parameters: the initial learning rate of the first to fifth convolutional layers is set to 0.0001; the initial learning rate of the parameters of the fully connected layer is 0.001; 40% reduction;

(4-3)加载训练数据，所述训练数据包括训练集和验证集；采用总共约有25万张图像的AVA大规模数据集对网络模型进行训练，对应网络的输入大小，所有样本统一归一化到256*256的大小。每张图片具有两个高、低质量两个标签中的一个；(4-3) Load the training data, the training data includes the training set and the verification set; the AVA large-scale data set with about 250,000 images in total is used to train the network model, corresponding to the input size of the network, and all samples are unified One to the size of 256*256. Each image has one of two high and low quality labels;

(4-4)采用随即梯度下降算法对初始化后的图像质量测试模型进行迭代训练，在训练集上，每迭代1000次保存一次模型参数，经过不断迭代，取得网络最优解，取在验证集上误差最小的模型作为已训练的图像质量评估模型。(4-4) Use the random gradient descent algorithm to iteratively train the initialized image quality test model. On the training set, save the model parameters every 1000 iterations. After continuous iterations, obtain the optimal solution of the network and take it in the verification set. The model with the smallest error above is used as the trained image quality assessment model.

(5)对目标图像使用已训练的质量评估模型进行测试。本发明的评价模型在AVA测试集上的分类准确率达到76.94％。(5) Test the target image using the trained quality assessment model. The classification accuracy rate of the evaluation model of the present invention on the AVA test set reaches 76.94%.

上述实施例为本发明较佳的实施方式，但本发明的实施方式并不受所述实施例的限制，其他的任何未背离本发明的精神实质与原理下所作的改变、修饰、替代、组合、简化，均应为等效的置换方式，都包含在本发明的保护范围之内。The above-mentioned embodiment is a preferred embodiment of the present invention, but the embodiment of the present invention is not limited by the embodiment, and any other changes, modifications, substitutions and combinations made without departing from the spirit and principle of the present invention , simplification, all should be equivalent replacement methods, and are all included in the protection scope of the present invention.