Channel DropBlock: An Improved Regularization Method for Fine-Grained Visual Classification

Yifeng Ding, Shuwei Dong, Yujun Tong, Zhanyu Ma, Bo Xiao, Haibin Ling

Introduction

Fine-Grained Visual Classification (FGVC) represents a significant challenge in computer vision, requiring models to distinguish between highly similar categories, such as specific bird species or car models. The primary difficulty lies in low inter-class variance—where different species look nearly identical—and high intra-class variance—where members of the same species vary significantly in pose or lighting.

Standard deep learning models often suffer from "co-adaptation," where they over-rely on a single dominant discriminative feature (e.g., a bird's beak) while ignoring other subtle cues. To address this, we present Channel DropBlock (CDB), a lightweight, "plug-and-play" regularization technique designed to prevent co-adaptation and encourage the discovery of multiple visual cues.

Methodology

Channel DropBlock (CDB) operates on feature maps within a convolutional neural network during training. Unlike standard Dropout, which masks individual units, or spatial DropBlock, which masks contiguous regions, CDB identifies and masks groups of correlated channels. This effectively "blinds" the model to specific visual patterns, forcing it to seek alternative discriminative features.

Overview of the Channel DropBlock (CDB) architecture, showing the process from correlation metric calculation to mask application.

The CDB process involves four main steps during training:

  1. Correlation Calculation: Determining which channels respond to similar visual patterns using either a Max Activation Metric (spatial proximity of peaks) or a Bilinear Pooling Metric (mathematical similarity of feature maps).
  2. Random Selection: Choosing a seed channel at random to initiate the drop process.
  3. Group Masking: Identifying the top $\gamma$ (a hyperparameter) most correlated channels to the seed and masking the entire group out.
  4. Normalization: Scaling the remaining features to maintain the signal's magnitude.

Experimental Results

The effectiveness of CDB was evaluated on standard FGVC benchmarks: CUB-200-2011 (birds), Stanford Cars (CAR), and FGVC-Aircraft (AIR).

Performance Gains

CDB consistently improved performance across various architectures. For instance, adding CDB to a ResNet50 baseline on the CUB dataset increased accuracy from 86.6% to 87.2%. When integrated into state-of-the-art models like PMG, it achieved a top accuracy of 89.9%. Notably, the method showed significant gains on CIFAR-100 compared to CIFAR-10, highlighting its specific utility for fine-grained tasks where subtle differences matter most.

Visual Evidence

Grad-CAM visualizations demonstrate that models trained with CDB develop a more holistic view of objects. While baseline models often focus on a single discriminative region, CDB-enhanced models distribute their attention across multiple parts (e.g., head, body, and tail of a bird), proving that the model is forced to learn a more diverse set of features.

Grad-CAM heatmaps comparing the baseline method with CDB variants, illustrating how CDB forces the model to focus on multiple discriminative regions.

Conclusion

Channel DropBlock provides an efficient way to improve FGVC performance without adding any inference-time overhead. By breaking co-adaptation through correlated channel masking, it forces networks to learn more robust and diverse feature representations. Future work will focus on automating hyperparameter tuning and exploring adaptive grouping strategies to further enhance the method's versatility across different datasets.

Summary updated

BibTeX
@inproceedings{ding2021channel,
  author = {Yifeng Ding and Shuwei Dong and Yujun Tong and Zhanyu Ma and Bo Xiao and Haibin Ling},
  title = {{Channel DropBlock: An Improved Regularization Method for Fine-Grained Visual Classification}},
  booktitle = {BMVC},
  year = {2021},
  url = {https://openreview.net/forum?id=r7oCEM59ql7},
}
Figure