Research And Application Of Image Complexity Based On Convolution Neural Networks

Influenced by the ambiguity of the definition of complexity and the difficulty of quantitative expression,there is no unified concept of image complexity evaluation method so far.Different research fields give different definitions and representation methods of image complexity.This paper tries to give the definition,representation and evaluation method of image complexity,and through the proposed complexity evaluation strategy to evaluate the complexity of image classification,image retrieval and face recognition data sets,and then give reasonable explanations and inferences to the evaluation indicators using statistical knowledge,hoping to provide a reasonable reference standard for future image processing research.The main research contents are as follows:Firstly,based on Monte Carlo Method and Ising Model,two kinds of data sets with 100*100 and 224*224 specifications are generated for image complexity evaluation,and relevant statistical descriptions and evaluations are carried out to validate the availability of the data sets.Then,on the basis of five groups of 10 traditional deep learning convolutional neural networks,the parameters are adjusted,the training and testing functions are improved,and the network model is trained,tested and saved.The results show that the verification output loss of a single network model in the verification set is very small,almost negligible.The statistics of the verification output of the same data set on multiple network models are all negligible.The difference of quantities is very small and the frequency distribution function approximates the same.This shows that this subject can evaluate the image complexity with the help of convolution neural network model,and the evaluation results are good.Next,this paper proposes a CADNet network model,which designs the DenseNet Network Model of Cyclic Control and Attention Mechanism for cyclic control and attention mechanism.The original intention is that the deep structure of the network model enlarges the flow of information and helps to improve the difficulty of network training.Although DenseNet network model structure has made efforts in this respect,it is believed that In addition to DenseNet's original traffic,CADNet model structure adds attention mechanism and features output to all layers in front of each layer,so that the number of input features obtained by each layer is twice as much as DenseNet's.The experimental results on verification set show that with a small amount of training intensity,CADNet model can be compared.Previously,the network worked better.Finally,the complexity of image classification,image retrieval and face recognition data sets is evaluated by using CADNet network,and the statistical description of the evaluation results in the data concentration trend,dispersion degree and distribution form is made.Finally,an objective conclusion is drawn.The experimental results show that,on the one hand,all data sets show negative skewness on the skewness features of the distribution shape,on the other hand,the peak features of the data distribution show a sharp,fat tail shape,and the trend of centralization within the group is obvious;on the other hand,the discrete degree of the image detection data groups is the largest,followed by the image classification data sets,and the performance of the face recognition data sets is basically the same.