Classification Of Fabric Patterns Image Based On Modified AlexNet

Image data increases explosively along with the rapid advancement of Internet technology.How to efficiently and accurately select the required image in the massive fabric image database has posed certain challenges to the image classification technology.At present,it mainly adopts manual annotation characteristics methods or traditional digital image processing classification method based on color and texture in the classification of fabric patterns.These classification methods extremely rely on a large amount of professional domain knowledge and practical experience,which limits the development on the classification of fabric patterns to some extent.Therefore,to find a kind of more robust and efficient classification method in fabric pattern classification task is still very urgent.Aiming at the problem of low classification efficiency and poor accuracy of manual classification in fabric pattern classification task,a classification model based on modified AlexNet has been proposed.The optimization is mainly based on the following two aspects.The first is the optimization of network structure,and the second is the optimization of parameters.In the first aspect,aiming at the “dead” problem caused by ReLu,which forces zeroing in the process of back propagation when updating parameter,the improved activation function Log-ReLu is proposed.It solves the neuron “dead” problem while retaining the fast convergence characteristic.Then for the massive amount of training parameters and the large required storage space in the process,a two-dimensional hash coded hidden layer is added on the network model.It makes the data be efficiently stored and calculated,and the classification more efficiency.In the second aspect,for the problem that the initial value of the learning rate heavily depends on the empirical value,a better learning rate initialization scheme and adaptive learning rate algorithm is proposed to further optimize the network performance and improve the classification accuracy.To solve the gradient dispersion problem that may be caused by Gaussian small random initialization,ASFABP algorithm which combined the artificial fish swarm algorithm and the BP algorithm is proposed.The method is to obtain the optimal weight initialization of neural units on the hash coded and the f8 layers.It optimizes the deep net performance and makes the network converge faster.Finally,performance is evaluated on Log-AlexNet,AlexNetNet,Vgg-16 and ResNet with the fabric pattern image dataset.Our numerical experiments show that the improved Log-AlexNet algorithm is superior to the original AlexNet on the fabric patterns image databases.It has higher classification accuracy and faster convergence speed.It's a better model to handle classification task.