Research On Image Open Set Recognition Method Based On Feature Fusion And Adversarial Learning

Deep learning has achieved advanced performance in image recognition,but most image recognition algorithms learn under limited label data conditions,and actual testing tasks require identifying targets in open large amounts of data.Therefore,identifying the openness of a dataset is more difficult than identifying a set of closed data.It requires not only a model to determine the classification accuracy of known categories,but also a model to determine unknown categories.Most existing methods have achieved good results,but in actual scenarios,there are intra distribution situations where known and unknown classes have similar potential feature representations,which gradually reveals the drawbacks of some open set recognition methods.This article proposes an image open set recognition method based on features fusion and adversarial learning,introducing supervised adversarial learning method and combining extreme value models to identify known and unknown classes in open datasets.The main contributions are summarized as follows:(1)This paper proposes an image open set recognition method based on multi-layer feature fusion to address the issue of insufficient robustness in open set recognition due to the imperfect distribution of potential feature spaces in known class samples.On the basis of the end-to-end network between the encoder and decoder,the feature extraction fusion methods are improved.In the encoder feature extraction stage,the deeper three-layer features in the backbone network are extracted,and a branch network is constructed to further extract the three-layer features.The decoder decodes and fuses different features of the network layer,improving the potential spatial distribution of the target to a certain extent,Finally,unknown classes are identified using decoding errors and extreme value models;In the aspect of known class classification,a feature correction training strategy based on parameter sharing is proposed to correct the multi-level features input to the decoder at the same time using the classification loss function,so that the features output by the encoder can not only improve the recognition accuracy of the algorithm for known classes,but also not affect the recognition of unknown classes.(2)This paper proposes an adversarial learning based open set recognition method for the distribution of urinary sediment images,which addresses the problem of image distribution with similar known cells and unknown impurities,as well as the slow convergence speed of model training.Gaussian radial basis function decoding network is used to make the model approach the decoding loss function,which speeds up the convergence speed of network learning;Then,use the anti loss function to learn the decoding error,separate the known class from the unknown class in the decoding error space,and optimize a large number of samples close to the Decision boundary through the singular value balanced weighting processing,and finally obtain a better separation effect through the extreme value model.(3)This paper conducted multiple experiments on multiple commonly used image classification datasets to validate the image open set recognition method of multi-layer feature fusion,and applied it to the urine sediment image distribution internal open set recognition task to validate the image distribution internal open set recognition method based on adversarial learning.In the open set state,necessary cells for disease diagnosis are classified,and impurities with similar potential feature distribution to effective cells are uniformly recognized as unknown classes,And compare it with other open set recognition methods.The experimental results show that the performance of our method in image open set recognition is superior to other methods.