.png){kind=icon}
.png){kind=icon}
Return
Enhancing Pruning Efficiency via Maximum Prunable Channel Selection in Tiny Object Detection
.png){kind=icon}
PRE
AI
DOI：10.1007/978-981-95-3052-6_14.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Tiny object detection presents significant challenges in computer vision due to inherent limitations such as low-resolution features and insufficient contextual information. While model pruning has emerged as a critical technique for compressing deep neural networks, existing sparse regularization methods often fail to achieve high pruning rates while preserving detection accuracy, especially in YOLO-based architectures. This paper proposes a novel Maximum Prunable Channel Selection (MPCS) algorithm to address the sparsity in pruning YOLOv12 models for tiny object detection. First, we analyze the constraints of conventional sparsity-inducing methods and identify that fixed regularization strategies result in suboptimal channel-wise sparsity distributions. To address this, a cosine-annealed gamma scheduling mechanism is introduced to dynamically adjust the sparsity rate, enabling balanced layer-wise regularization. Subsequently, the MPCS algorithm selectively identifies and eliminates redundant channels with maximal prunability scores, thereby decoupling sparsity constraints from the pruning process. Experiments on the benchmark TinyPerson and Visdrone datasets demonstrate that, compared to conventional methods, our approach achieves a 10% improvement in model pruning ratio while maintaining stable detection accuracy. The proposed approach offers a practical solution for deploying lightweight yet accurate detectors in resource-constrained scenarios.
Keywords:
Channel Pruning
Model Compression
Channel Selection Algorithm
YOLOv12
Journal
K
IF:
0
Papers:
30
Citations:
0