Image Feature Extraction And Application Based On Deep Network

In the processing of images,how to describe the image features becomes a key problem because of the variety of types of images and the complexity of the information conveyed by a single image.The quality of the information delivered by the extracted image features will directly affect the results of the image processing.Traditional hand-designed image features are cumbersome and inefficient.The widely used neural network in recent years has solved this problem very well.The neural network uses a learning-based image feature extraction method.This method uses image data as the input of a neural network,transforms it through a series of stacked linear combinations and nonlinear transformations into a higher level abstract representation,and automatically extracts image feature information layer by layer.Depth networks with multiple hidden layers can often make a deeper and more in-depth description of the original input data,thereby learning more advanced data feature representations.For the research of image feature extraction based on deep network,this paper mainly studies the following contents:Two kinds of depth image feature extraction tools are studied in depth: stacked sparse autoencoder and convolutional neural network,and its structure and training process are analyzed.And it is applied to feature extraction in two image processing tasks: image fusion and image recognition: lung nodule detection and false positives.Stacked Sparse Autoencoder(SSAE)is an efficient unsupervised feature extraction method with good complex data representation capabilities.Different from traditional deep learning methods requiring a lot of labeled data.Stacked sparse autoencoders are unsupervised deep learning methods that are suitable for image fusion tasks.By adding sparse constraints in the hidden layer of the SSAE,the sparse features of the image can be obtained as an accurate representation of the source image,effectively expressing the internal structural features of the image.We combines an advanced multiscale decomposition tool,the translation invariant shear transform(SIST),and proposes a new image fusion method based on SSAE extraction of image features.First,the source image is decomposed into low-frequency subbands and high-frequency subbands by SIST.Second,two-layer SSAE is used as a feature extraction method to obtain deep sparse representations of high frequency subbands.Then,the fusion of maximum SSAE-based features is proposed.The rule is to fuse the high-frequency sub-band coefficients;then the weighted average fusion rule is used to merge the low-frequency sub-band coefficients;finally,the fusion image is obtained by inverse SIST transform.The experimental results show that this method is superior to the traditional method in subjective and objective evaluation.Convolutional neural network(CNN)uses local connections to efficiently extract image features.Using a three-dimensional network structure can effectively express three-dimensional spatial feature information in sequence images.This feature makes it suitable for feature extraction in medical image processing.False positive reduction is a key procedure of computer-aided pulmonary nodule detection.The objective of it is to recognize the true pulmonary nodule from the plentiful candidates which received from the first step of pulmonary nodule detection.The difficulty in getting to the false positives lies in how to extract effective feature expressions from the 3-dimensional lung image sequence to distinguish between true lung nodules and some other highly deceptive lung tissues.Recently residual network is more and more popular around the world with it's distinguished performance.A multi-context three-dimensional residual convolutional neural network(3D Res-CNN)was proposed to realize the reduction of the false positive nodule.The proposed algorithm uses two scales of networks to adapt to changes in lung nodule size,and uses residual connections to reduce network errors and learn more characteristic image features.In addition,in order to alleviate the imbalance of data,firstly,a small number of categories of original image blocks are rotated and resampled;then different weights are assigned to different categories of data when calculating the cost function.Experiment on volumetric computed tomography(CT)data indicates that our method gets state of the art performance: 0.843 average sensitivity with 0.125,0.25,0.5,1,2,4,8 false positive per subject.The result declared the effectiveness of residual convolutional network for the recognition of the true pulmonary nodule from the plentiful candidates.In summary,this paper applies two kinds of image feature extraction methods based on deep network to two image processing tasks respectively,effectively expressing the essential feature information of the image.The experimental results show that the feature extraction method based on deep network has more powerful image information representation capability.