Research On Enhancement And Recognition Of Mongolian Furniture Pattern Based On Multi-Scale Convolutional MSRCR

The Mongolian furniture pattern is formed by the Mongolian people after thousands of years of cultural accumulation.It contains the unique thoughts and feelings of the Mongolian people and is a typical symbol of multi-ethnic integration.Mongolian furniture has a wide variety of patterns and rich color composition.It is a unique manifestation of Chinese traditional culture and needs to be inherited and carried forward.In the process of collection and recognition of Mongolian furniture patterns,due to the influence of factors such as environment,equipment,human interference and the weathering of patterns,low contrast,abnormal brightness,blurred patterns,color distortion and degradation will occur,which will affect the subsequent recognition and classification.In response to these problems,this thesis combines digital image processing technology and convolutional neural network to study three types of Mongolian furniture patterns of animals,geometry and plants,and realizes the enhancement and recognition of Mongolian furniture patterns.The main research contents of this thesis are as follows :(1)The principles and enhancement methods of Gamma correction and SSR,MSR and MSRCR algorithms based on Retinex theory are introduced.MSR is an algorithm to improve the detail and color of pattern image,but it is easy to produce halo and overexposure as SSR algorithm.MSR algorithm based on particle swarm adaptive Gamma correction uses PSO algorithm to optimize gamma parameters and adjust brightness,which can reduce the occurrence of halo and overexposure.The improved MSR algorithm is superior to the MSR algorithm,but the problems of color loss and color degradation are not solved.In this thesis,a MSRCR algorithm based on multi-scale convolution is proposed.The MSRCR algorithm is improved by steps such as secondary guided filtering,multi-channel multi-scale convolution and linear weighted fusion,and white balance.The proposed algorithm is subjectively compared with other algorithms.The results show that the algorithm improves the contrast and brightness of the pattern image,improves the details,maintains the color fidelity,and weakens the color degradation.(2)The enhancement algorithm proposed in this thesis is compared with GC algorithm,SSR algorithm,improved MSR algorithm and MSRCR algorithm.The subjective evaluation and peak signal-to-noise ratio(PSNR),mean square error(MSE),structural similarity(SSIM),average gradient(AG)and entropy(E)are used to analyze the enhanced Mongolian furniture patterns.The results show that the proposed algorithm can achieve the best enhancement effect compared with other algorithms.(3)The pattern and background of the pattern image are segmented by the improved Kmeans clustering segmentation algorithm.The algorithm quickly determines the value of the clustering number K by the number of gray histogram peaks and uses the Mahalanobis distance instead of the Euclidean distance to solve the random fluctuation problem.(4)The Res Net-34 model in Convolutional Neural Network(CNN)is used to identify animal,geometric and plant Mongolian furniture patterns.The structure and principle of the recognition model are described.The recognition results obtained by different enhancement algorithms and whether they are segmented are evaluated by the recognition rate and the accuracy,recall,precision and F1 scores calculated by the confusion matrix.The results show that the Mongolian furniture pattern image is first enhanced by the multi-scale convolution MSRCR algorithm and segmented by the improved K-means algorithm,and then the recognition effect obtained by the Res Net-34 model recognition is the best.The optimal loss value of the training data set is about 0.225,and the overall recognition rate of the Mongolian furniture pattern is about 86.90 %.