Adversarial Auto-encoder For Open Set Recognition

Open set recognition is a typical task in computer vision.Compared with the image recognition task under the closed set hypothesis,the open set recognition task can not only identify the known categories in the training data set,but also make a rejection response or label as "unknown" the unknown category objects that do not appear in the training data set.In the real world recognition or classification task,due to the limitation of various objective factors,it is usually difficult to collect training samples of all categories when training the recognizer or classifier.More realistic situation is Open Set Recognition(OSR),training the existence of incomplete category knowledge of the world,when testing allows the unknown class instance is submitted to the algorithm,for classifier not only accurate class of existing in the process of training model of classification,but also effectively dealing with unknown classes in the training data set.With the rise of deep learning in recent years,deep learning for image recognition depends on convolution model to a large extent.While Convolutional Neural Networks(CNNs)are effective,they are rarely explained afterwards.The most recent method to solve the open set recognition task consists of two main parts: introduce distance metric on the objects belonging to the same category and make them have lower inner-class distance;Then a compact decay probability model is established.For testing an object instance,when the probability is lower than the threshold value of all known classes,it is determined that the object belongs to a category not seen in the training dataset.However,these methods do not take into account the label information of the known categories of image data in the training process of the model.On the basis of previous studies,the label information of known categories of image data is incorporated into the training phase of the model.Specifically,the main contributions of this paper are three:(1)According to the requirement of the number of auto-encoder network parameters in specific experimental environment,the number of network channels and batch processing size were adjusted.By controlling variables,the performance of the auto-encoder network with different number of network channels and batch processing size is tested,and compared with the experimental results of open set recognition,the performance of the auto-encoder in the open set recognition model can be maximized.(2)Because the research trend from closed set recognition to open set recognition is just in its infancy,previous researchers have ignored the image category label information in training data.In this paper,by embedding the label information of training data in the image reconstruction,the geometric features obtained by different categories of images on the deep network can be larger margins,and the inner-class distance between image features of the same category can be more compact.The discriminability of inter-class image instances and the cohesion of inner-class image instances are improved.This allows better identification of known categories in the training dataset.At the same time,the class instances of the open set space that do not appear in the training data set can be make a rejection.(3)In order to verify the validity of tag information of embedded image data and the generalization and applicability of the open set recognition model in this paper,extensive experiments are carried out on several standard data sets.Experimental results show that this method is better than the existing deep open set classifier in many standard data sets,and has strong robustness to different open set categories.