ABSTRACT The core challenge of fine‐grained image recognition (FGIR) tasks is distinguishing highly similar subclasses within the same base category. Most CNN‐based deep learning methods typically focus on extracting information from local regions while overlook the inherent structure between subclasses and the complex relationships between features. This paper presents a two‐stage feature selection method based on partial order analysis (POA) and heterogeneity evaluation (HE) for FGIR tasks, guiding the model to focus on distinctive features while reducing uncertainty caused by interfering information. Specifically, in the POA stage, clustering first groups similar subcategories into a medium‐granularity category. Formal concept analysis then models their hierarchical partial order, identifying “shared features” among subcategories and “exclusive features” unique to each. This structured representation highlights key contrastive cues. In the HE stage, a novel heterogeneity index is introduced to measure the fluctuation of low‐level features within each fine‐grained category. This index guides the model to suppress pseudo‐discriminative features with high heterogeneity, mitigating the impact of noisy and unstable information on decision‐making. We perform comprehensive experiments on three commonly used benchmark datasets (CUB‐200‐2011, Stanford Cars, and FGVC‐Aircraft). Experimental results show that the proposed method outperforms classic FGIC methods, validating the effectiveness of our approach.

Stacking pseudoinverse learner (SP) is an ensemble learning technology, and its generalization performance greatly affects the effect of image classification. Currently, most SPs randomly initialize the input weight matrix in a random subspace without limiting the random initial values, resulting in unstable training results and a decrease in generalization performance; in addition, training all samples at once may cause the classifier redundant and also affect the generalization performance of the model. To efficiently address the above issues, we propose a new framework called progressive stacking pseudoinverse learner (PSP), which aims to enhance the generalization performance of SP via active learning (AL) in random subspaces. Specifically, on the one hand, a random feature SP (RFSP) model is proposed, which constrains the random subspace by initializing the input weight matrix into different random specific distributions to improve the generalization performance of SP. On the other hand, an AL progressive (ALP) model based on RFSP is proposed. By iteratively selecting useful samples to optimize the classification results, the training sample information is effectively used to progressively enhance the generalization performance of the model. Experimental results on three public datasets show that our proposed PSP algorithm achieves better performance in accuracy, precision, recall, and$F1$score, and the results are competitive with state-of-the-art methods.

In zero-shot learning (ZSL) community, it is generally recognized that transductive learning performs better than inductive one as the unseen-class samples are also used in its training stage. How to generate pseudo labels for unseen-class samples and how to use such usually noisy pseudo labels are two critical issues in transductive learning. In this work, we introduce an iterative co-training framework which contains two different base ZSL models and an exchanging module. At each iteration, the two different ZSL models are co-trained to separately predict pseudo labels for the unseen-class samples, and the exchanging module exchanges the predicted pseudo labels, then the exchanged pseudo-labeled samples are added into the training sets for the next iteration. By such, our framework can gradually boost the ZSL performance by fully exploiting the potential complementarity of the two models' classification capabilities. In addition, our co-training framework is also applied to the generalized ZSL (GZSL), in which a semantic-guided OOD detector is proposed to pick out the most likely unseen-class samples before class-level classification to alleviate the bias problem in GZSL. Extensive experiments on three benchmarks show that our proposed methods could significantly outperform about 31 state-of-the-art ones.

In this paper, we introduce a method for building LOD (levels of detail) modeling from urban multi-view stereo (MVS) meshes. Using city MVS meshes as input, our algorithm proceeds in three main steps: segmentation, contour extraction and modeling. With the prior knowledge and span constraint, we first segment the scene with an adapted variational measure to discover the underlying structures. The next contour extraction step projects the vertical structures onto the ground as line segments and extract the facade contours from them with a Markov random field. In the last modeling step, the contours are used to label the roof sections out and extruded to generate models of LODs with semantics. Experiments on complex and noisy urban meshes show that our approach could generate compact and accurate building models when compared with stateof- art methods.