Data scarcity in base classes restricts the diversity of learned semantic features, leading to poor generalization and unclear class boundaries. To enhance feature robustness, we apply targeted visual transformations, focusing on color and shape as suggested by previous studies. Specifically, we employ random cropping to expand the fine-grained feature space. Given an image X, the cropping dimensions are defined as wcrop∼Rand(wmin,wmax) and hcrop∼Rand(hmin,hmax). Given a rand point ( xr, yr), where xr∈[0, w−wcrop], yr∈[0, h−hcrop]. Thus, the point ( xf, yf) of the lower right corner of the final cropping area are computed in Equation 1:

Random cropping of varying sizes is employed to capture local information, enhancing both local feature understanding and fine-grained semantic perception. To further enrich class-aware semantics, we introduce a transformation set ℱ∈{Tc,Tr}, consisting of color jittering ( Tc) and random rotation ( Tr). The processed RGB images are then transformed into the frequency domain using the 2D DCT, which expresses pixel data via a linear combination of cosine basis functions. Leveraging the superior energy compaction of DCT over the complex-valued DFT (He et al., 2020; Huang et al., 2025), each channel of the input image X is converted to the frequency spectrum P2d in Equation 2:

where h∈{0,1,…H−1}, w∈{0,1,…W−1} represent the horizontal and vertical frequency indices. The normalization coefficients αh and αw are defined as in Equation 3:

In the resulting spectrum P2d, low-frequency components are concentrated near the origin (0,0), while high-frequency components are distributed in the peripheral regions. Then, we apply a binary mask M to separate the spectrum into low-frequency and high-frequency components. We define a cut-off threshold τ based on the Manhattan distance in the frequency domain. The mask M is defined as in Equation 4:

Subsequently, the low-frequency spectrum Plow(2d) and high-frequency spectrum Phigh(2d) are derived via the Hadamard product ( ⊙): Plow(2d)=P(2d)⊙M,Phigh(2d)=P(2d)⊙(1−M). Finally, we project the masked spectra back to the spatial domain using the 2D Inverse DCT (IDCT) in Equation 5:

These components are then sent into the encoder and obtain the original image IX, low-frequency Il and high-frequency Ih feature maps. This enables the extraction of discriminative details from high-frequency components and structural context from low-frequency components. The feature extraction is defined as in Equation 6:

Utilizing enhanced discriminative feature maps as prior knowledge augments the model’s ability to capture critical information and adapt to incremental data. Specifically, by aligning samples with class prototypes via the high-frequency features Ih′, we encode fine-grained details that effectively sharpen decision boundaries and enhance model performance. Then, the embedded image is computed by Equation 7:

Where C0 is designated as the encountered class. This can extend various semantics and fill the unallocated image embedding space. It can also provide semantic knowledge, which encourages extensive learning of different semantics for better generalization.

For label domain, the predefined alternations can generate multiple augment image-label pair ( x, y), where ℱ(x, y)={xn, yn}n=1N. N is the number of transformed extension space. xn are the generated extension images, and the corresponding labels is yn=y×N+n. Thus, the label space is extended with the fine-grained class-aware embedding derived from the original space. The association between the image domain and the label domain can effectively provide richer semantic details to improve the accuracy. Likewise, the training within the embedding space ℱ can be expressed by Equation 8:

Although effective for coarse classification, existing methods are limited in handling fine-grained data. We therefore propose an embedding-based supervised contrastive learning strategy using the MoCo (He et al., 2020) framework. This method optimizes feature distances by clustering positive pairs and separating negative ones. Given an instance (x,y), we generate a query view xq=Augq(x) and a key view xk=Augk(x) via data augmentation. A shared encoder fq, comprising a feature extractor and a classifier, is then employed to extract the corresponding features. As shown in Equation 9:

where ωTϵℝd×|C0| is the weight value, and f(x)ϵℝd×1 is the feature function. Among them, the query encoder fq are encoded through gradient descent, while the key encoder fk are encoded by a progressing encoder, driven by a momentum update with the fq. The queue of key embedding is maintained to store the feature vector.

In the label domain, a label queue maintains labels corresponding to the feature queue, facilitating the differentiation of positive and negative pairs. This queue preserves an identical length to the feature queue. Subsequently, the contrastive loss is computed to drive the model to capture discriminative fine-grained features. This optimization effectively minimizes intra-class distance while maximizing inter-class variation, thereby fostering deep interaction between the visual and label domains.

With the incremental sequence, the backbone network is fixed, and the classifier is computed by computing the novel class prototypes. The novel class information can be acquired to enable the extension of classifiers with the prototypes of basic classes and extended-augmentation classes. As shown in Equation 14:

where b is the number of basic classes, N is the number of transformed extension space, and t is the number of incremental sequences. The prototypes Wn represents the focus of global and local fine-grained semantics from the original classes, shown in Equation 15:

Subsequently, the classifier is updated by the novel classes’ prototypes combined with the original class prototypes. It can help push the novel samples away from the distributions of the old classes and benefit novel class-aware semantic information generalization. The FC layer of the model is updated by contrasting novel query samples with these slowly evolving key embedding of base classes from the feature queue. Finally, the cosine similarity between embedding and all prototypes is computed to obtain the inference results for the test image. It can be formulated as Equations 16 and 17:

By adapting the classifier to the novel classes while keeping the backbone unchanged, it can maximize the preservation of previously acquired knowledge.

We split our dataset into the training set and the testing set. Specifically, 75% of the data is used for training, and the remaining 25% is used for testing. The basic training phase is composed of 16 classes with 200 samples. And there are 4 incremental sequences, within 3 classes and 5 samples of each class. The experimental results of each class are detailed in Table 1.

As shown in Table 1, the identification performance is notably strong for the base classes. In particular, the model demonstrates high accuracy and robustness in classifying classes A, E, C, F, D, H, and I. In contrast, the F1-scores for the incremental classes predominantly range between 0.3 and 0.6, highlighting the model’s limited effectiveness in learning new classes. For instance, classes R and Q exhibit high recall but low precision; this can be attributed to significant inter-class feature similarity, which leads to misclassification. To further evaluate our model’s performance, we calculated the confusion matrix, and the experimental results are depicted in Figure 7.

While the identification results for base classes are comparable across methods, the improvements in incremental classes are more pronounced. In the confusion matrices, a brighter diagonal indicates higher identification accuracy, with significant contrast areas highlighted in red and green. As shown in Figure 7A, for base class identification, the marked areas demonstrate that FGDE achieves superior performance with fewer misclassifications compared to other methods. When comparing Figure 7A with Figure 7B, our model demonstrates distinct advantages over the baseline lacking high-frequency and low-frequency enhancement, particularly in the highlighted regions. Mechanistically, low-frequency components capture global structural features, while high-frequency components extract local fine-grained details, thereby enriching the image representation. Conversely, while the method without frequency-aware extension maintains a visible diagonal for base classes, it performs poorly on novel classes. In contrast, our method exhibits robust performance, indicating that it effectively adapts to novel classes without disrupting previous decision boundaries.

To better evaluate the performance of the model and enhance the explanation of training, the loss and accuracy results of our model are shown in Figure 8.

The loss curves and their convergence trends are illustrated in Figure 8. Throughout the training process, the loss consistently decreased while accuracy improved, eventually leading to model convergence. Specifically, the trajectories of the multi-objective losses are detailed in Figure 8A. The loss exhibits a steady decline until stabilizing at approximately 80 epochs, with the model achieving peak accuracy at epoch 88. Evaluations on the testing set confirm the model’s robust classification performance.

Simultaneously, as illustrated in Figure 9, While the CE loss baseline provides marginal class separation, our proposed method demonstrates superior capability in distinguishing base classes and integrating novel classes with minimal feature overlap.

To verify the effectiveness of our model, we visualize the identification results using a scatter plot, as shown in Figure 10. The horizontal axis represents the True Labels, while the vertical axis denotes the Predicted Labels. In this visualization, points aligned closely with the diagonal indicate accurate classification performance. Conversely, points deviating from the diagonal represent misclassifications, where the magnitude of deviation highlights the discrepancy between the predicted and ground truth labels.

In Figure 10A, the points are predominantly clustered along the diagonal, indicating that the model achieves robust overall classification performance. In contrast, while Figure 10B exhibits some alignment with the diagonal, a significantly larger number of points deviate from it. This dispersion is particularly pronounced in specific categories, such as classes 7 and 8, indicating higher misclassification rates. For the newly added classes, Figure 10A maintains relatively high accuracy despite occasional errors. Conversely, Figure 10B reveals a marked decline in performance for these later classes, evidenced by a substantial increase in misclassified points and greater deviations from the diagonal. In summary, our method leverages a frequency-based separation strategy: low-frequency components extract fundamental structural features, while high-frequency components capture fine-grained details. This approach enhances inter-class separability and minimizes the interference of new classes on existing representations. Consequently, our method demonstrates significant advantages, exhibiting superior accuracy and robustness.

To evaluate the accuracy performance of our model, the state-of-the-art FSCIL models are compared with ours. The experimental results are shown in Table 2.

Our proposed method achieves a peak accuracy of 95.000% on the base classes, surpassing all existing baselines. Across subsequent incremental sessions, our method maintains robust performance, recording accuracies of 91.701% in the first session and 78.906% in the fourth. Notably, catastrophic forgetting is significantly better mitigated compared to competing approaches.

While most methods suffer performance degradation as new classes are introduced, distinct patterns emerge. iCaRL exhibits the most inferior performance, starting with a base accuracy of 70.542% and declining rapidly, resulting in the lowest Harmonic Mean (HM) of 58.774%. Although FSLL, FACT, and C-FSCIL achieve respectable base accuracies (91.421%, 92.040%, and 94.051%, respectively), they experience sharp drops in later sessions, yielding HM scores below 80%. SAVC and Wang’s method demonstrate relatively stronger resilience, with HMs of 84.328% and 83.101%, respectively. Nevertheless, our approach consistently outperforms these leading methods across all sessions, securing the highest HM of 86.599%.

This superior performance is attributed to our frequency decomposition strategy. By separating images into low-frequency components (capturing global structural features) and high-frequency components (preserving fine-grained details), we generate a detail-enhanced discriminative representation. This mechanism bolsters both base class separability and novel class generalization. In summary, our model retains base class knowledge while adapting to new sequences with minimal degradation, marking a significant advancement over state-of-the-art methods in solving the FSCIL challenge.

Class Activation Maps (CAMs) are essential for interpreting model decisions by highlighting influential image regions. For each instance, the original image is shown with its corresponding CAMs in Figure 11. The color intensity represents the activation level, signifying the importance of each region in the model’s classification result.

From Figure 11, a systematic comparison across both base and incremental classes reveals a consistent pattern of superior performance by our method. Our approach demonstrates significantly more accurate localization of target objects, with activations that adhere tightly to object boundaries while effectively suppressing background noise. For instance, in items (1), (4), (8), and (10) (Wang et al., 2024), produces diffuse activations that often spill into the background or focus on restricted, peripheral regions. In contrast, our method generates heatmaps that are precisely centered on the targets, covering their salient regions more effectively. Furthermore, our model consistently yields more comprehensive activation maps that encompass the entire object, suggesting the acquisition of a holistic representation. This is particularly evident in items (2), (5), (7), and (11). Whereas (Wang et al., 2024) tends to fixate on local textures or edges, our model captures the full semantic structure of the object. This holistic understanding is crucial for robust classification, rendering the model less susceptible to variations in orientation or partial occlusion. Moreover, the heatmaps generated by our method exhibit a more concentrated focus on the objects’ discriminative regions. In contrast, the activations in the Wang (Wang et al., 2024) model appears scattered and less intense, as evident in examples (5), (9), and (12). Our model, conversely, produces strong, focused activations localized on the core features of the objects. This indicates that our approach more effectively identifies key predictive features and is less prone to relying on spurious image correlations. These qualitative results strongly support our hypothesis that the proposed methodology facilitates learning more robust and interpretable feature representations. By generating more complete and accurately localized heatmaps, our model demonstrates a deeper semantic understanding of the image content and enhanced mitigation of catastrophic forgetting.

We compare our method with other state-of-the-art methods on the Chinese Medicine dataset, and the comparison results are shown in Figure 12. From Figure 12A, the base classes are 8, the number of classes in each incremental subsequent is 4, and the N-way is set to 3. We can see that our model exhibits the highest accuracy among other mainstream methods. From Figure 12B, the model is initialized with 60 base classes, followed by incremental sessions containing 8 classes each, with the setting of 3-shot samples per class.

As observed in Figure 12, iCaRL exhibits the most precipitous performance decline, suggesting that its replay-based strategy lacks the stability required for FSCIL tasks and struggles to mitigate interference from novel classes. FSLL attains high initial accuracy but suffers a sharp drop in later stages, indicating that despite early effectiveness, its long-term generalization capability is limited. FACT and C-FSCIL display a more gradual decay, reflecting stronger knowledge retention, although performance degradation persists. Conversely, SAVC and the method by Wang et al. maintain relatively stable performance, particularly in later sessions. Furthermore, to evaluate the statistical reliability of our results, we add the error bars to represent the standard deviation in Figure 12. Our method is consistently compact for incremental sessions in both Figures (A) and (B). This low variance indicates that our FGDE model is highly robust to initialization differences and data sampling fluctuations, maintaining stable performance even as the number of classes increases. In contrast, baseline methods such as iCaRL and FSLL exhibit larger error bars in several sessions (e.g., Session 0 and 3 in Figure 12B), suggesting higher instability. Notably, our method outperforms all competing approaches, achieving consistent improvement and minimal degradation. By leveraging fine-grained feature comparison and multi-semantic discrimination, our approach significantly enhances adaptability to novel categories while effectively mitigating catastrophic forgetting.

To select the appropriate backbone to extract features in our model, the various backbones are compared to reflect the impact of model efficiency. We also evaluate the model with a hybrid transformer structure, such as MBConv (Howard et al., 2017).

Using running time as a key feasibility metric, we compared various backbones in Table 3. ResNet20 and ResNet50 showed lower accuracy and longer inference times compared to ResNet18. VGG’s simpler architecture limited its feature extraction, reducing base class accuracy, while EfficientNet and DenseNet121 proved computationally expensive due to their complexity. Results from MBConv further highlighted the limitations of hybrid transformer structures in this setting. Although VGG19 improved incremental class accuracy by 1.322%, its slower running time made it less viable. Thus, ResNet18 was selected for achieving the highest comprehensive performance.

To evaluate the learning ability and generalization ability of our model to incremental classes, experiments with different numbers of classes are performed for the basic training. This paper compares the base 12 classes with the 16 classes. When the base classes are 12, the number of classes in each incremental subsequent is 4, and N-way is set to 4. When the base classes are 16, the number of classes in each incremental subsequent is 3, and N-way is set to 3. The experimental results are shown in Table 4.

For base 12 classes, the base accuracy is a better 1.875% than the 16 classes. However, base 12 classes perform 10.380% worse than base 16 classes in the incremental sequences. The experimental results demonstrate that a greater number of base classes enhances the ability of the model to learn and capture fine-grained features more effectively. Finally, we set the base class to 16, N-way is 3.

In the original settings, images are initially cropped to enrich the fine-grained feature space and increase image diversity. To evaluate the impact of cropping sizes on model performance, our experiment examines the different cropping sizes. The comparison results are shown in Table 5.

From Table 5, it is evident that crop size significantly affects feature diversity. The 128-crop size achieves the highest discrimination, surpassing the 96-crop and 32-crop settings by 7.909% and 11.980%. Additionally, the similar HM scores for crop sizes 96 and 64 suggest a consistent feature distribution at these scales. Given these results, we fixed the crop size at 128 to ensure optimal model performance.

To verify the effectiveness of our networks for few-shot images, we also compared the impact of different numbers of base classes on model performance. To ensure fairness, the remaining parameters remain unchanged in the comparison. The experimental results are shown in Table 6.

The results indicate that the accuracy for base classes remains relatively stable regardless of the class count. However, regarding the identification of incremental classes, the configuration with 200 base classes yields an accuracy 8.536% higher than that with 250 base classes and 5.667% higher than that with 300 base classes. These findings suggest that 200 base classes offer the optimal balance between feature diversity and model generalization.