طبقه بندي تصاوير بافتي نويزدار با استفاده از شبكه عصبي عميق و الگوي دودويي م حلي كامل

Noisy Textures Classification Using Deep Neural Network and Completed Local Binary Pattern

اﻟﮕﻮي دودوﯾﯽ ﻣﺤﻠﯽ ﯾﮏ ﺗﻮﺻﯿﻔﮕﺮ ﭘﺮﮐﺎرﺑﺮد در اﺳﺘﺨﺮاج وﯾﮋﮔـﯽ از ﺗﺼـﺎوﯾﺮ ﺑـﺎﻓﺘﯽ اﺳـﺖ. ﺷـﺒﮑﻪﻫـﺎي ﻋﺼـﺒﯽ ﻋﻤﯿـﻖ ﮐﺎﻧﻮﻟﻮﺷـﻦ ﻧﯿـﺰ از ﺑﻬﺘﺮﯾﻦ اﺑﺰار ﻃﺒﻘﻪ ﺑﻨﺪي ﺑﺎ دﻗﺖ ﺑﺴﯿﺎر ﺑﺎﻻﺑﻪ ﺷﻤﺎر ﻣﯽ رود. ﻧﻮآوري اول اﯾـﻦ ﭘـﮋوﻫﺶ اراﺋـﻪ ﯾـﮏ ﺳـﺎﺧﺘﺎر ﺑـﺮاي ﺗﺮﮐﯿـﺐ وﯾﮋﮔﯽ ﻫـﺎي اﻟﮕـﻮي دودوﯾﯽ ﻣﺤﻠﯽ و ﺷﺒﮑﻪ ﻋﺼﺒﯽ ﻋﻤﯿﻖ ﮐﺎﻧﻮﻟﻮﺷﻦ ﺑﺮاي ﻃﺒﻘﻪﺑﻨﺪي ﺗﺼﺎوﯾﺮ ﺑـﺎﻓﺘﯽ ﻧـﻮﯾﺰدار اﺳـﺖ ﮐـﻪ دﻗـﺖ ﺑﺴـﯿﺎر ﺑـﺎﻻﯾﯽ را ﺑـﺮاي ﻃﺒﻘﻪ ﺑﻨـﺪي ﺗﺼﺎوﯾﺮ ﺑﺎﻓﺘﯽ ﻧﻮﯾﺰدار ﻓﺮاﻫﻢ ﻣﯽ ﮐﻨﺪ. اﯾﻦ روش از دو اﺑﺰار اﺳﺘﺨﺮاج وﯾﮋﮔﯽ ﺗﺸﮑﯿﻞ ﺷﺪه اﺳﺖ. در ﯾﮏ اﺑﺰار، ﺑﺎ اﺳﺘﻔﺎده از اﻟﮕﻮي دودوﯾـﯽ ﻣﺤﻠﯽ ﮐﺎﻣﻞ وﯾﮋﮔﯽﻫﺎي ﻣﺤﻠﯽ ﺗﺼﺎوﯾﺮ ﺑﺎﻓﺘﯽ، در ﻗﺎﻟﺐ ﻫﯿﺴﺘﻮﮔﺮام ﺳﻪ ﺑﻌﺪي اﺳﺘﺨﺮاج ﻣﯽﺷﻮد. در اﺑﺰار دوم وﯾﮋﮔﯽﻫﺎي ﺑـﺎﻓﺘﯽ ﺑـﺎ اﺳـﺘﻔﺎده از ﺷﺒﮑﻪ ﻋﺼﺒﯽ ﻋﻤﯿﻖ ﮐﺎﻧﻮﻟﻮﺷﻦ DenseNet-121 ﮐﺎﻫﺶ داده ﻣﯽ ﺷﻮد. اﯾﻦ ﺑﺨﺶ ﮐﻪ در ﻓﺮآﯾﻨﺪ ﺗﺮﮐﯿﺐ وﯾﮋﮔﯽ ﻫﺎ ﺑﮑﺎرﮔﺮﻓﺘـﻪ ﺷـﺪه اﺳـﺖ، ﺑﺎ اﺳﺘﻔﺎده از ﺷﺒﮑﻪ ﻋﺼﺒﯽ ﮐﺎﻧﻮﻟﻮﺷﻦ ﮐﻢ ﻋﻤﻖ، اﺑﻌﺎد ﻫﯿﺴﺘﻮﮔﺮام ﺳﻪ ﺑﻌﺪي را ، ﺑﺮاي ﺗﺮﮐﯿﺐ ﺑﺎ وﯾﮋﮔﯽ ﻫﺎي ﻋﻤﯿﻖ، ﺑﻪ ﺻﻮرت ﻗﺎﺑـﻞ ﺗـﻮﺟﻬﯽ ﮐﺎﻫﺶ ﻣﯽدﻫﺪ. دﻗﺖ ﻣﺪل ﭘﯿﺸﻨﻬﺎدي، روي ﻣﺠﻤﻮﻋﻪﻫﺎي داده ﻧﻮﯾﺰي CUReT ،Outex و UIUC ﺑﺎ ﻧﻮﯾﺰ ﮔﻮﺳـﯽ، ﻧـﻮﯾﺰ ﻧﻘﻄـﻪ اي و ﻧـﻮﯾﺰ ﻓﻠﻔﻞ ﻧﻤﮑﯽ ﺑﺎ ﺷﺪﺗﻬﺎي ﻣﺨﺘﻠﻒ، ارزﯾﺎﺑﯽ ﺷﺪه اﺳﺖ و دﻗﺖ ﻃﺒﻘﻪﺑﻨﺪي روش ﭘﯿﺸﻨﻬﺎدي ﺑﺮاي ﻣﻘﺎدﯾﺮ ﻣﺨﺘﻠـﻒ ﻧـﻮﯾﺰ، ﺑﻬﺒـﻮدي ﺑـﯿﻦ 3 ﺗـﺎ 15 درﺻﺪ داﺷﺘﻪ اﺳﺖ.

Local binary pattern is one of the most popular descriptor that widely used in feature extraction of texture images. Deep convolutional neural network is also one of the best classification methodthat provides very high accuracy. In this research, by combining the features that produced by these two methods, a structure for noisy texture classification is proposed, which provides a very high classification rate. This method is based on two extracted features. The first part uses completed local binary pattern features and in the second part the features of texture images are extracted by using the DenseNet-121 convolution deep neural network. Another motivation of this research related to feature reduction, which significantly reduces the dimensions of extracted features. It employs a shallow convolution neural network to convert the extracted features into lower number of new features. The accuracy of the proposed method has been evaluated on noisy Outex, CUReT and UIUC datasets. The classification accuracy of the proposed method for different level of noise has increased significantly compared to many advanced methods and has improved between 3 and 25%.