BDNet improves image classification with fast and slow feature layers
BDNet: Fast-slow thinking is all you need
Computer Vision and Pattern Recognition
Summary
Image classification systems usually compare parts of an image directly to each other, which can be slow or use lots of computing power. The authors present BDNet, which instead compares image features to slow-changing reference statistics gathered during training. This approach speeds up processing and keeps computations simple while aiming to improve accuracy and robustness to image noise or lighting changes. Their tests use different versions of BDNet on the ImageNet dataset and check its performance under normal and disturbed conditions.
What this means in practice
- •For mobile app developers: Integrate efficient image classifiers that handle noisy inputs rapidly to improve app performance on constrained devices.
- •For security system engineers: Deploy compact and robust image recognition models for surveillance cameras that maintain accuracy under varying light and visual noise.
Authors
Chenzhi Zheng, Zhuoran Zheng
Abstract
We present BDNet, a hierarchical image-classification backbone built around fast image features and slow statistical references. Its baseline deformation operator (BDO) transforms the deviation of a feature from a stage-specific reference through a low-rank projection and a learned channel scale. Rather than computing pairwise token similarity, the branch operates directly on baseline-relative residuals with complexity linear in the token count. Each of four reference buffers accumulates detached stage statistics during training and remains frozen at inference. A four-stage pyramid with depths $(2,2,6,2)$ and widths $(96,192,384,768)$ yields a 20.70M-parameter classifier at $224\times224$ resolution. We organize an ImageNet-1K study around clean classification, eight controlled ablations, and Gaussian-noise, contrast, and brightness perturbations. The central hypothesis is that a slow reference can improve the robustness of a compact representation while retaining a simple tensor implementation. We distinguish this training-time statistical mechanism from input-dependent attention and make its linear inference behavior explicit.