Research On Deep Convolution Image Recognition Method Based On Data Mining

With the development of the Internet and the explosive growth of informat io n,people often face to processing the huge data with complex interferences.Traditio na l methods are usually difficult to get satisfaction results for such big data.Instead,deep learning has shown tremendous superiority.Deep Convolutional Neural Networks(CNN)are widely used to recognize a large number of images with complex backgrounds.But the computational speed of traditional CNNs are usually slow due to the large amount of calculations.In order to reduce this shortcoming,some researches have been studied in this dissertation.The specific work is as follows:(1)Segment interest area from the image.In order to reduce the interference of background to image recognition,one target segmentation methods are proposed.It is to build a CNN to segment the interest area based on statistical methods.(2)Establish some deep CNN SAR image recognition models to analyze the factors affecting the recognize results.First CNN models with different topologies are trained by the segmented regions of interest targets.Then the influence of the number of convolutional layers and channel in each convolutional layer,and the size of the convolution kernel are analyzed to the final recognition result.At last,the topologica l structure of the deep CNN recognition model can be determined based on the analysis results.(3)Establish a multi-channel deep CNN image recognition model.Obviously the size of convolution kernel determines the image regions of the local features extracted by the deep CNN model.The existing deep CNN platforms have same size for convolution kernels in the same convolution layer,so the final local features with different geometric sizes cannot be extracted by a deep CNN.However,it is likely that local features with different sizes are needed to obtain higher correct recognition rate in practical applications.In order to fully capture the local features of different scales of the target,a multi-path convolution based on convolution kernels of different sizes is established in this paper.Network model to improve the target's correct recognit io n rate.(4)Screen convolution features based on data mining.The features extracted by the multi-channel convolutional neural network model have very large redundancy.The dispersion ratio and the weighted contribution rate of the fully connected layer neurons are used to screen the features extracted by the multi-channel network.Then the features are further refined by genetic algorithms.The removal of redundant features not only improves the correct recognition rate of the target,but also greatly improves the recognition speed.If the network needs to learn again,it can also greatly increase the speed of relearning and reduce the training samples required for relearning.In order to improve the accuracy and speed of target recognition,a convolutio na l neural network is established firstly to segment the interest image areas combined with statistical methods.Then a multi-channel deep CNN recognition model is established and a series of data mining methods are used to screen the features extracted by deep CNN to optimize the topologies structure of the deep CNN.Finally,the MSTAR database was used to verify the proposed method experimentally.Experiments show that an effective convolution segmentation model can be quickly established based on statistical methods,the multi-channel deep convolution image recognition model can improve the accuracy of image recognition and data mining methods can effective ly optimize the structure of the multi-channel CNN image recognition model,which can not only improve the recognition accuracy,but also the recognition speed.