Enhancing Fine-Grained Classification for Low Resolution Images

Maneet Singh, Shruti Nagpal, Mayank Vatsa, Richa Singh

Abstract

Low resolution fine-grained classification has widespread applicability for applications where data is captured at a distance such as surveillance and mobile photography. While fine-grained classification with high resolution images has received significant attention, limited attention has been given to low resolution images. These images suffer from the inherent challenge of limited information content and the absence of fine details useful for sub-category classification. This results in low inter-class variations across samples of visually similar classes. In order to address these challenges, this research proposes a novel attribute-assisted loss, which utilizes ancillary information to learn discriminative features for classification. The proposed loss function enables a model to learn class-specific discriminative features, while incorporating attribute-level separability. Evaluation is performed on multiple datasets with different models, for four resolutions varying from 32x32 to 224x224. Different experiments demonstrate the efficacy of the proposed attribute-assisted loss for low resolution fine-grained classification.

The Challenge of Low-Resolution Fine-Grained Classification

Fine-grained classification involves distinguishing between highly similar sub-categories, such as specific animal breeds or vehicle models. While high-resolution imagery allows deep learning models to identify subtle discriminative features, performance degrades significantly as resolution decreases. In low-resolution scenarios—typical in long-range surveillance or mobile photography—essential details like textures or specific anatomical shapes vanish, making visually similar classes like a Wolf and a German Shepherd nearly indistinguishable.

Comparison of high-resolution (128x128) and low-resolution (32x32) images showing the loss of fine details.

Attribute-Assisted Loss Methodology

To address the information deficit in low-resolution images, this research proposes a novel Attribute-Assisted Loss. This method leverages ancillary attribute data (e.g., "spotted," "domestic," "forest-dwelling") during the training phase to supervise the structure of the learned feature space.

The total loss function combines standard Cross-Entropy for classification with a specialized Attribute Loss ($L_{Attribute}$). The intuition is to minimize the distance between an image's feature representation and the mean feature vector of all images sharing the same attribute. This forces the model to group images by shared attributes while simultaneously maintaining class-level separation. Crucially, these attribute labels are only required during training; at inference, the model predicts class labels directly from low-resolution inputs using the optimized feature space.

The proposed Attribute-Assisted Loss framework, showing how images are mapped to a feature space organized by both class and attribute.

Experimental Evaluation

The approach was evaluated on two major datasets across resolutions ranging from 32x32 to 224x224:

  • Animals with Attributes-2 (AwA-2): 37,322 images across 50 animal categories, utilizing 85 binary attributes (e.g., white, domestic, patches).
  • DeepFashion: Over 800,000 images of clothing using category-level attributes like upper-body, lower-body, and full-body.

The experiments utilized standard architectures including ResNet-18, ResNet-50, and DenseNet-121 to demonstrate the model-agnostic nature of the loss function.

Key Results and Insights

The proposed loss consistently outperformed standard training methods, with the most dramatic gains observed at the lowest resolutions. For instance, on the AwA-2 dataset at 32x32 resolution, the method achieved 53.24% accuracy using ResNet-50, a significant jump from the 38.09% achieved with standard Cross-Entropy loss.

Key findings include:

  • Attributes as Regularizers: In small datasets or low-information regimes, the attribute-assisted loss acts as a regularizer, preventing the model from overfitting on blurred pixels.
  • Non-Visual Cues: The model benefits from both visual (color, pattern) and conceptual (habitat, temperament) attributes, suggesting that high-level semantic grouping aids visual classification.
  • Resolution Sensitivity: While the benefit is most pronounced at 32x32, the method still provides a performance edge at higher resolutions (e.g., 93.73% at 224x224 on AwA-2 vs. 86.79% for baseline methods).

Conclusion

The Attribute-Assisted Loss provides a robust framework for enhancing fine-grained classification when visual details are scarce. By incorporating semantic attributes into the feature learning process, models can maintain discriminative power even at extreme low resolutions. Future work will explore handling multiple simultaneous attributes and evaluating performance on real-world low-resolution data containing sensor noise and motion blur.

Summary updated

Abstract

Low resolution fine-grained classification has widespread applicability for applications where data is captured at a distance such as surveillance and mobile photography. While fine-grained classification with high resolution images has received significant attention, limited attention has been given to low resolution images. These images suffer from the inherent challenge of limited information content and the absence of fine details useful for sub-category classification. This results in low inter-class variations across samples of visually similar classes. In order to address these challenges, this research proposes a novel attribute-assisted loss, which utilizes ancillary information to learn discriminative features for classification. The proposed loss function enables a model to learn class-specific discriminative features, while incorporating attribute-level separability. Evaluation is performed on multiple datasets with different models, for four resolutions varying from 32x32 to 224x224. Different experiments demonstrate the efficacy of the proposed attributeassisted loss for low resolution fine-grained classification.
BibTeX
@inproceedings{singh2021enhancing,
  author = {Maneet Singh and Shruti Nagpal and Mayank Vatsa and Richa Singh},
  title = {{Enhancing Fine-Grained Classification for Low Resolution Images}},
  booktitle = {IJCNN},
  year = {2021},
  url = {https://openreview.net/forum?id=8g2NReOutlux},
}
Figure