Research On Methods Of Image Classification Based On Capsule Architecture

Image classification is one of basic and important tasks of computer vision.Driven by deep learning and application requirements,image classification methods have been greatly developed in recent years.The classification objects have expanded from grayscale images,RGB images to SAR images,hyper-spectral images,medical images and so on.The task scenario changes from a closed and static environment with sufficient samples to an open and active environment with limited samples.Deep convolutional neural network is a powerful tool for image classification tasks,but it has structural defects such as excessive attention to texture features and insensitivity to the spatial location information of entities in the image,and it shows limitations when dealing with classification tasks in an open environment.In order to overcome these shortcomings,the capsule network came into being.It replaced the scalar output of the neuron with the vector output of the capsule.It showed advantages in exploring the spatial position of the entity and extracting contour information.However,it also has high algorithm complexity and bad performance when handling complex background images.In fact,due to the variety of imaging mechanisms and task requirements,it is very necessary to design a reasonable neural network structure for specific task characteristics.Therefore,we synthetically combine the convolutional network and the capsule network to design three effective solutions for the classification problems in three specific open environments:· We propose a capsule-based SAR image classification method for SAR target recognition task in complex environment.Although deep convolutional neural networks have been widely used in SAR target recognition tasks,the classification models based on deep convolutional neural networks always perform poorly on handling the SAR images,which are occluded,noisy,or the observation elevation angle changes greatly.To this end,we propose a novel SAR capsule network to make full use of the spatial relationship between different entities in SAR images,and achieve accurate and robust classification of SAR images,without increasing computational complexity.Specifically,we design a shallow convolutional network for feature extraction,and then conduct the classification through a classifier based on the capsule architecture.The SAR capsules learn through a vector-based fully connected operation instead of a traditional routing process,which not only reduces thecalculation cost,but also improves the recognition accuracy.In evaluations with rotation,occlusion,additive noise and multiplicative noise,the SAR capsule network exhibits stronger robustness than the typical convolutional neural network.When missing training samples in a certain azimuth range or existing a large depression variance between training samples and test samples,the proposed network achieves better performance than the existing works and reveals some competitive advantages in several evaluation scenarios.· We propose a conditional variational auto-encoder based on the capsule architecture for open set recognition.In open set recognition,a classifier has to classify known class samples as well as detect unknown class samples.In order to accurately identify unknown classes,the classifier needs to project samples of the known classes into a compact and separable region in the feature space to discriminate samples of unknown classes.The capsule network is capable of learning strong representative feature.The components of the capsule represent the different attributes of the image entities,which is suitable for the open set recognition problem.During training,we propose a multi-prior distribution strategy to encourage different known class capsules to match different prior distributions.To this end,we use the conditional variational auto-encoder framework to force the posterior distribution of the encoder output to approximate a set of Gaussian prior distributions.In this way,we can control the compactness of the features belonging to the same class around the center of the Gaussians,thus improving the ability of unknown detection of the classifier.We conducted several experiments and obtain state of the art results on different datasets in the open set recognition and unknown detection tasks.· For solving few-shot open-set recognition,we propose an attention mechanism based on affine transformation and an open space management strategy based on contrast learning.Most of the existing few-shot recognition methods are designed for closed-set scenarios and trained by meta-learning methods,thus how to perform few-shot recognition in an open environment has not been fully explored.Considering the advantages of the capsule network in learning representative features,we draw the inspiration from Transformer to simplify the capsule architecture into a plug-and-play attention mechanism,which can facilitate the metric-based metalearning model to adapt to the open environment.In addition,we utilize the proto- types to formulate the open space,and propose a contrast loss term to manage the open space.Experiments show that these two strategies can effectively improve the performance of the metric-based meta-learning model on few-shot open-set recognition tasks.