Discriminant Learning Based Open Set Recognition Using Feature Ensemble Strategy

Open Set Recognition(OSR)is different from common classification tasks.It requires the model to correctly classify samples of known categories and to identify samples of unknown categories.It extends classification problems from closed world assumptions to open scenarios.Open set recognition research has great significance for safety-related applications such as autonomous driving and fault diagnosis.Unknown class samples can be easily misclassified as known classes because of the general classification models’ single feature space with close inter-class distance.Based on the models,OSR systems are difficult to achieve as good results as closed set systems.In addition,most existing work only evaluates the performance of open sets on standard datasets.However,unlike standard datasets,the category distribution in real scenarios is not only open but also long-tailed,meaning it has category imbalance,which affects the model’s robustness and reduces its OSR performance.In order to improve the recognition performance of OSR in real scenarios,this work proposes a network framework that integrates and optimizes feature representations,and constructs a feature representation that is discriminative and robust to class imbalance.In terms of feature integration,because the feature representation of a single feature space has local activation characteristics,the classifier is prone to generate high confidence for unknown class samples.Therefore,this work builds a set of parallel networks with the same structure,learns multiple representations through attention perception from different transformations of image data,diversifies the feature representations and proposes a corresponding collective decision-making method.In terms of feature representation optimization,this work discriminatively optimizes the feature representation of a single feature space in a parallel network,and proposes a Combined Euclidean and Angular Space Clustering Loss(EACL)function.This loss function makes the samples form tight clusters at the centers of their categories,achieving compactness within categories and large distances between categories,and provides a global unknown space for unknown category recognition.To address the long-tail characteristics of category distribution in real scenarios,this work proposes a C-Softmax(Center Softmax)loss framework and applies it to EACL.It makes the network discriminative enough for the open set and robust to class imbalance.This work builds an open-set recognition algorithm that integrates multiple feature spaces and optimizes feature representations in a single space to be effective in real-world scenarios.The proposed method achieves state-of-the-art open-set detection performance on multiple standard object recognition datasets and outperforms the state-of-the-art methods by 2.6% on the most challenging standard dataset Tiny Image Net with an openness of57.36%.At the same time,it surpasses the existing work in plankton identification of real scene datasets with openness ranging from 8.6% to 40.24%,achieving an average of 3.9%higher performance under different openness levels.