Search

Fine-grained visual classification (FGVC) is challenging due to the subtle inter-class variations. Key region location and discriminative feature extraction are the crucial aspects of FGVC. In this paper, we carry out research from the perspective of multi-granularity and multi-scale, explore the information in different granularities and scales, and propose a novel method called Multi-Scale Network via Progressive Multi-Granularity Attention (MSMGA-Net), which locates the key discriminative regions while avoiding recognition confusion caused by subtle inter-class variations. In order to achieve discriminative region locating, we adopt a progressive training strategy to make the network work in steps during training process, and the jigsaw puzzle generator (JPGen) is constructed to generate images with different granularity for different training steps. The focus of each step of training is to locate the key discriminative regions of specific granularity image at the corresponding stage of the network through the multi-granularity attention (MGA) module. Aiming to obtain the discriminative features, we also construct a multi-scale convolution (MSC) module for feature extraction corresponding to each step respectively, enabling the capture of different scales of details. We validated our MSMGA-Net through extensive experiments on CUB-200-2011, FGVC Aircraft and Stanford Cars. The experiments demonstrate that MSMGA-Net can get promising results. In particular, the accuracy on two publicly competitive benchmark datasets, CUB-200-2011 and FGVC Aircraft, reached 90.30% and 93.84%, respectively. • A framework is designed that incorporates attention and multi-granularity feature. • A multi-granularity attention module is proposed to locate key regions. • A multi-scale convolution module is constructed to extract multi-scale features. • A jigsaw puzzle generator is constructed to generate different granularity images. [ABSTRACT FROM AUTHOR]

Understanding of geoscience relies on spatio-temporal continuous fields, such as Sea Surface Height (SSH). Data reconstruction for restoring spatio-temporal continuous and gridded maps from incomplete observations and inaccuracy interpolation products has long been a crucial challenge in the marine geoscience. Despite remarkable progress, most existing learning-based reconstruction methods neglect to fully and discriminatively utilize the inter-layer and intra-layer features. The inter-layer semantics are complementary, and the significance of intra-layer components varies with frequency. To address these issues, we propose a multi-layer Feature Combination Network based on Attention mechanism (FCANet) for SSH data reconstruction. Specifically, a novel trainable Multi-layer Feature Combination Block (MFCB) is developed to enrich the features through inter-layer dependencies. Dual attention mechanism is introduced into the MFCB to enhance saliency of intra-layer features by adaptively rescaling the spatial-wise and channel-wise features. Furthermore, experimental results on SSH variable demonstrate the superiority of our FCANet over state-of-the-art reconstruction methods in the metrics of RMSE and SSIM. [ABSTRACT FROM AUTHOR]

Learning subtle and discriminative regions plays an important role in fine-grained image recognition, and attention mechanisms have shown great potential in such tasks. Recent research mainly focuses on employing the attention mechanism to locate key discriminative regions and learn salient features, whilst ignoring imperceptible complementary features and the causal relationship between prediction results and attention. To address the above issues, we propose an Attention Interaction and Counterfactual Attention Network (AICA-Net). Specifically, we propose an Attention Interaction Fusion Module (AIFM) to model the negative correlation between the attention map channels to locate the complementary features, and fuse the complementary features and key discriminative features to generate richer fine-grained features. Simultaneously, an Enhanced Counterfactual Attention Module (ECAM) is proposed to generate a counterfactual attention map. By comparing the impact of the learned attention map and the counterfactual attention map on the final prediction results, quantifying the quality of attention drives the network to learn more effective attention. Extensive experiments on CUB-200-2011, FGVC-Aircraft and Stanford Cars datasets have shown that our AICA-Net can get outstanding results. In particular, it achieves 90.83% and 95.87% accuracy on two open competitive benchmark datasets CUB-200-2011 and Stanford Cars, respectively. Experiments demonstrate that our method outperforms state-of-the-art solutions. [ABSTRACT FROM AUTHOR]