Abstract Image sample augmentation refers to strategies for increasing sample size by modifying current data or synthesizing new data based on existing data. This technique is of vital significance in enhancing the performance of downstream learning tasks in widespread small-sample scenarios. In recent years, GAN-based image augmentation methods have gained significant attention and research focus. They have achieved remarkable generation results on large-scale datasets. However, their performance tends to be unsatisfactory when applied to datasets with limited samples. Therefore, this paper proposes a semantic similarity-based small-sample image augmentation method named SSGAN. Firstly, a relatively shallow pyramid-structured GAN-based backbone network was designed, aiming to enhance the model’s feature extraction capabilities to adapt to small sample sizes. Secondly, a feature selection module based on high-dimensional semantics was designed to optimize the loss function, thereby improving the model’s learning capacity. Lastly, extensive comparative experiments and comprehensive ablation experiments were carried out on the “Flower” and “Animal” datasets. The results indicate that the proposed method outperforms other classical GANs methods in well-established evaluation metrics such as FID and IS, with improvements of 18.6 and 1.4, respectively. The dataset augmented by SSGAN significantly enhances the performance of the classifier, achieving a 2.2% accuracy improvement compared to the best-known method. Furthermore, SSGAN demonstrates excellent generalization and robustness.

Pattern recognition from two-dimensional (2D) images is an indispensable tache in computer vision. However, one legacy hinders its progress: multifarious visual distortions (e.g., partially occluded signs, fisheye respective, affine or 3D projections) caused by spatiotemporal-varying perturbations (e.g., shrunk, sharpening, overexposure, jitter, and motion) significantly degrade the performance of neural networks in terms of high-level intelligent behaviors (e.g., target localization and recognition). Leveraging the growing availability of capsule network (CapsNet), we suppress the deformation effect to the final prediction by proposing a CapsNetORB framework to implement distorted target recognition. Two highlights, the customized Siamese CapsNet (S-CapsNet) and vector-based oriented fast and rotated brief (VB-ORB), can cast a mutual positive stimulation: the former encodes capsule feature vectors for the later, whilst the later detects space-scale invariant interval dimensions (instead of pixels) to bridge association between source standard images (high-quality training images) and distorted ones (testing images). Thus, the category of one source standard image owning the most correspondences is the final predicted category. We believe that capsule vectors own higher representability and stability compared with conventional pixels/feature maps, which can be well exploited in feature learning while resisting visual distortions. Experimentally, we show that employing our pipeline for distorted target categorization can outperform state-of-the-arts by delivering promising performance on CUB-200-2011, Stanford Dogs, Stanford Cars, and our hand-crafted data set.

The automatic topic labeling model aims at generating a sound, interpretable, and meaningful topic label that is used to interpret an LDA-style discovered topic, intending to reduce the cognitive load of end-users while browsing or investigating the topics. In this study, we first introduced the pre-trained language model BERT to topic labeling tasks. It exploits the contextual embedding of the pre-trained language model to improve the quality of encoding sentences. To generate a topic label with higher Relevance, Coverage, and Discrimination, we propose a novel summarization neural framework. Specifically, it exploits the paired-attention to model the relationship between the candidate sentences first and then decides which sentences should be included in the final summarization topic label. Moreover, we expected that high-quality sentence encoding representation could improve our model's performance. So, for each discovered topic, we trained a specific layer to extract the important topic-related features from the sentence embeddings as well as filter the noise information. The experimental results showed that our model significantly outperforms the state-of-the-art and classic topic labeling models.