Exploiting Position Information in Convolutional Kernels for Structural Re-parameterization

Tianxiang Hao, Hui Chen, Guiguang DingIJCAI 2025 · DOI

Abstract

Position Boosting Convolution (PBConv) is a novel structural re-parameterization technique that enhances CNN performance by adaptively emphasizing the most informative spatial positions within a convolutional kernel. By training with a set of parallel small kernels and merging them into a single standard convolution for inference, PBConv delivers significant accuracy gains across classification, segmentation, and detection tasks with zero additional computational cost at runtime.

Introduction

Structural Re-parameterization has emerged as a powerful paradigm in deep learning, allowing models to benefit from complex, over-parameterized architectures during training while maintaining efficiency during deployment. Existing methods like ACNet or RepVGG typically focus on adding external branches (e.g., $1 \times 1$ convolutions). However, they often overlook the internal spatial properties of the standard $K \times K$ kernel.

Ordinary convolution baseline: identical structure during training and inference.

The authors of this work observed that not all positions in a $3 \times 3$ kernel contribute equally to feature extraction. Depending on the task and layer depth, specific positions (such as the center or corners) may be significantly more informative. PBConv addresses this by explicitly "boosting" these critical positions during the training phase.

Methodology: Position Boosting Convolution (PBConv)

The core innovation, PBConv, replaces a standard convolution with a multi-branch structure during training.

1. Parallel Small Kernels

Instead of a single $3 \times 3$ convolution, PBConv employs several concurrent smaller kernels (e.g., $1 \times 3, 3 \times 1, 2 \times 2$) that collectively cover the spatial extent of the original kernel. This allows the network to learn more nuanced spatial features.

The PBConv framework: multi-branch training with parallel small kernels followed by equivalent conversion to a single kernel for inference.

To manage the vast search space of kernel combinations, the authors developed a fast heuristic search. Candidate structures are randomly generated and trained for a brief period (1-10% of total training time). The structure exhibiting the highest weight dispersion (standard deviation) is selected, as this signal reliably predicts final model accuracy.

3. Linear Merging for Inference

Since convolution and Batch Normalization (BN) are linear operations, the parallel branches can be mathematically fused into a single $3 \times 3$ weight matrix and bias vector post-training. This ensures that the inference model is structurally identical to a standard CNN, incurring no extra latency.

Experimental Results

PBConv was evaluated across diverse benchmarks, consistently outperforming baseline models and existing re-parameterization techniques like ACB and DBB.

  • Image Classification: On ImageNet, PBConv improved ResNet-18 top-1 accuracy by 1.33% and ResNet-101 by 1.13%.
  • Semantic Segmentation: For Cityscapes, the mIoU of PSPNet (ResNet-18) saw a 1.50% improvement.
  • Object Detection: CenterNet's mAP on COCO was boosted by 1.20%.
  • Image Deblurring: Gains of up to 1.05 dB PSNR were achieved on the GoPro dataset.

Qualitative analysis via Grad-CAM indicates that PBConv-equipped models focus more precisely on target objects while effectively suppressing background noise.

Key Insights

  • Magnitude and Importance: There is a strong correlation between the magnitude of a kernel weight at a specific position and the performance impact when that position is pruned.
  • Adaptive Positional Diversity: Importance maps vary by architecture; for instance, ResNet-18 favors the bottom-right, while VGG-16 shows a "center cross" pattern. This necessitates the adaptive nature of PBConv.
  • Orthogonality: PBConv can be combined with other re-parameterization methods (e.g., RepVGG) to achieve cumulative performance gains.

Conclusion and Future Work

PBConv provides a robust method for improving CNN performance without inference-time overhead. While highly effective for $3 \times 3$ kernels, future work may explore scaling the heuristic search to larger kernel sizes and further optimizing training efficiency to mitigate the slight increase in training-time complexity.

Summary updated

BibTeX
@inproceedings{hao2025exploiting,
  author = {Tianxiang Hao and Hui Chen and Guiguang Ding},
  title = {{Exploiting Position Information in Convolutional Kernels for Structural Re-parameterization}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/121},
}
Figure