Multi-scale Sparse Feature Learning For Complex Image Generation And Interpretation

In recent years,deep learning technology has made great progress in image processing,which has a wide range of applications,such as image classification,image segmentation,target recognition and tracking.The perception and characterization of image features are the key to image classification and subsequent image processing tasks.Therefore,it is of great significance to study effective feature extraction and representation.At present,the methods based on deep convolution neural networks have been widely used in image classification tasks.However,it lacks of interpretability and can not fully mine the internal structure information of images.As a real two-dimensional image representation method,multi-scale geometric analysis can capture the geometric structure of the image,such as edge,contour and texture.Among them,the contourlet transform has multi-scales,multidirections,anisotropy and good local characteristics,which can effectively describe the high anisotropy of the image.Aiming at the task of image classification,segmentation and reconstruction,this thesis studies the image processing methods based on multi-scale contourlet and CNNs by combining the convolution neural network in the spatial domain and the multiscale geometric analysis tool in the spectral domain,which is committed to improving the feature perception ability and interpretability of neural networks.The specific contents and contributions of this thesis are summarized as follows:(1)Aiming at the poor interpretability of convolution neural networks in image classification,a novel network architecture named contourlet sparse representation network is proposed,which creatively integrates the multi-resolution analysis and spatial convolution network in the same framework.Through the spatial-spectral feature fusion strategy,the multidirectional and multi-scale features can be effectively extracted in the proposed model.Thus the accuracy and robustness of the model are improved.The statistical feature fusion strategy increases the distinguishability between classes.The experimental results show that the model has achieved good classification results on both texture data sets and remote sensing data sets.It has certain advantages in the image classification task of small data sets.(2)Since the variability of appearance and the fuzziness of object boundaries in the image segmentation,a multi-scale contourlet knowledge guide learning network is proposed for accurate segmentation.By combining the sparse representation of the multi-scale contourlet with the parallel reverse attention network,the multi-scale distinguishable features are learned.In addition,the adaptive learning of contourlet guided features is realized through the contourlet knowledge adaptive learning.It captures region boundaries and segments small objects more accurately.Experiments show that the model has good performance in the widely used polyp data sets and remote sensing building extraction data sets,and the effectiveness of the method is verified on mixed datasets and cross datasets.(3)In order to enhance the biological interpretability of the convolutional neural network,inspired by the sparse and hierarchical features representation in the ventral stream of the human visual system,the biologically inspired multi-scale contourlet attention network(BMCAnet)is proposed.The multi-scale contourlet filter banks are used to construct a population of neurons in the primary visual cortex(V1),and extract sparse features in a multi-scale and multi-direction way.It simulated a simple cell in V1 that responds to stimuli in a specific direction.In order to refine contourlet features adaptively,the Shannon block attention module(SBAM)is introduced to learn the weights of contourlet coefficients adaptively.The proposed contourlet pooling layer can obtain the invariant structure features,which roughly stimulate the pooling process of complex cells in the V1 area.Experimental results show that the model can effectively extract sparse features and achieve good classification results on handwritten dataset,fashion product image dataset,texture datasets,and remote sensing scene datasets.(4)Most research on polarimetric SAR image classification is only conducted in the complex spatial domain and lacks effective representation in the complex spectral domain.A complex-valued contourlet neural network is proposed which integrates the multi-resolution analysis of nonsubsampled contourlet(NSCT)and enhances the feature representation.The corresponding weights and bias in the complex domain can be adaptively learned in an end-to-end fashion.To improve the distinguishability and classification ability of feature learning,the statistical feature integration module(SFIM)is proposed to capture the statistical properties of the NSCT coefficients.The competitive results demonstrate that the proposed model is better than the single spatial feature learning in the complex domain.Good classification results have been achieved on polarimetric SAR datasets.(5)Aiming at the challenge of over-smooth output caused by the existing image superresolution reconstruction network.A knowledge-based contourlet inference network is proposed to reconstruct the HR image through the series of corresponding contourlet coefficients.Specifically,first,we consider the low-pass subbands of contourlet as the corresponding LR image.Then,feed it to the embedding net with residual blocks to provide adequate information for the contourlet coefficients prediction.Finally,we innovatively convert the estimation of the contourlet coefficients into the estimation of the generalized Gaussian distribution(GGD)parameters,and design the corresponding loss function to ensure training stability,which explores the smoothness of the contour effectively and guarantees the general structure and details of images.Experiments on the remote sensing datasets demonstrate the accurate super-resolution reconstruction of the proposed model.