Multi-attention Meta Learning for Few-shot Fine-grained Image Recognition

Abstract

<jats:p>The goal of few-shot image recognition is to distinguish different categories with only one or a few training samples. Previous works of few-shot learning mainly work on general object images. And current solutions usually learn a global image representation from training tasks to adapt novel tasks. However, fine-gained categories are distinguished by subtle and local parts, which could not be captured by global representations effectively. This may hinder existing few-shot learning approaches from dealing with fine-gained categories well. In this work, we propose a multi-attention meta-learning (MattML) method for few-shot fine-grained image recognition (FSFGIR). Instead of using only base learner for general feature learning, the proposed meta-learning method uses attention mechanisms of the base learner and task learner to capture discriminative parts of images. The base learner is equipped with two convolutional block attention modules (CBAM) and a classifier. The two CBAM can learn diverse and informative parts. And the initial weights of classifier are attended by the task learner, which gives the classifier a task-related sensitive initialization. For adaptation, the gradient-based meta-learning approach is employed by updating the parameters of two CBAM and the attended classifier, which facilitates the updated base learner to adaptively focus on discriminative parts. We experimentally analyze the different components of our method, and experimental results on four benchmark datasets demonstrate the effectiveness and superiority of our method.</jats:p>

Selective, Structural, Subtle: Trilinear Spatial-Awareness for Few-Shot Fine-Grained Visual Recognition

“…Network architectures. Following the previous works [9], we adopt the standard feature extraction network Conv4. This contains four convolutional modules, each of which consists of a convolutional layer with 3×3 size followed by batch normalization layer and ReLU layer.…”

“…Specifically, for 5-way 1-shot task on benchmarks, we can increase by 17.81%, 14.37%, 29.22%, and 5.42% on average over Relation Net [1], DNN [3], Prototype Net [4], MattML [9]. For 5-way 5-shot task, we also can gain by 18.73%, 1.77%, 32.25%, and 2.89 % improvement with a large margin on average over Relation Net [1], DNN [3], Prototype Net [4], MattML [9]. The results mean that our approach enables the network to better extract the subtle discriminant features, which are a benefit for the similarity measure of fine-grained images.…”

“…Most of the existing works [1,2,3,4,5,6,7,8,9,10,11,12] in few-shot learning mainly consists of two groups: nearest neighbor based approaches [1,2,3,4,5,6,7] and attention mechanism based approaches [8,9,10,11,12]. The first group attempts to calculate the similarity scores of nearest neighbor samples by designing an effective distance function loss to optimize the model.…”

“…Wei et al [8] first tackle the few-shot fine-grained problem by jointing a piecewise mapping technique and bilinear feature learning module. Zhu et al [9] propose to utilize a multi-attention module to highlight the dominative information for few-shot fine-grain image classification. Despite their utilization of attention mechanisms, the contextual understanding of spatial relations, especially the multi-scale representation, is less explored.…”

Coupled Patch Similarity Network FOR One-Shot Fine-Grained Image Recognition

“…LRPABN [9] proposes a pairwise bilinear pooling operator to perform feature alignment for learning an efficient distance metric. To capture more discriminative features, Zhu et al [10] develop a multi-attention meta-learning method to learn informative parts by attention mechanism. Similarly, BSNet [11] is proposed to improve the generalization ability by decreasing the similarity bias and extracting more diverse features.…”