Research On The Color Measurement Model Based On Convolutional Neural NetWork And Processing Strategy Of Small And Medium Scale Samples

In the field of industrial printing,inspection of the color quality of printed products based on machine vision system has become an application trend.In the machine vision inspection system,the digital camera obtain images in the RGB color space.The RGB color space is device-dependent,and the RGB response of the same object in different imaging environment is different,color quality evaluation cannot be performed.Using a machine vision inspection system to obtain the true color of the object's surface,the RGB color space needs to be mapped into the device-independent CIEL*a*b*color space.In this thesis,multi-channel imaging data is used to propose a color space mapping method based on CNN(Convolutional Neural Network,CNN)network.At the same time,according to the color quality measurement standard for industrial printtings,the optimization strategy of the convolution neural network under small and medium scale samples is proposed.Three processing strategies are proposed to adjust the size of the network structure,introduce random forests,and transfer learning.In this thesis,the spectral reflectance reconstruction model based on CNN network is proposed to establish the mapping method from RGB color space to CIEL*a*b*color space for color measurement.First,multi-channel imaging data and their spectral reflectance are used to establish a mapping model of RGB color space to spectral reflectance based on a CNN network,and then the spectral reflectance is converted to a CIEL*a*b*color.Finally,conversion accuracy is evaluated using color difference and accuracy and consistency of spectral reflectance measures.The simulation results show that,based on the spectral reflectance reconstruction method,the CNN network model has higher measurement accuracy and less color aberration than the method based on traditional BP network.Therefore,the spectral reflectance reconstruction model based on the CNN network can better meet the requirements for the color quality inspection of printed materials.Aspects to the optimization problem of spectral reflection reconstruction model based on convolution neural network in small and medium scale samples,this thesis puts forward three kinds of processing strategies,such as adjusting convolution kernel size and network structure scale,designing convolution neural network-random forest hybrid model and transfer learning.The first is to verify the fitting performance of the network model by adjusting the convolution kernel size and the scale of network structure.The second is to use convolution neural network to extract the feature information of samples and provide it to the random forest for the establishment of model.The third is to use the transfer learning method to train a convolution neural network model using house price dataset,and then to train the model with multi-channel RGB response data.The simulation results show three kinds of processing strategies all have important influence,which can improve the color measurement precision of the model.