A GPU-based Framework for Adaptive Kernel Size Selection in Visual Pattern Recognition
Jiahong Wang, Zixuan Wang · International Journal of Pattern Recognition and Artificial Intelligence · 2025
Visual pattern recognition systems relying on fixed-kernel Convolutional Neural Networks (CNNs) often struggle with heterogeneous image regions, leading to suboptimal feature extraction. This paper introduces a GPU-based framework that addresses the challenges of adaptive kernel size selection, including parallelism disruption, resource contention, and real-time constraints. The framework comprises three key innovations: (1) an online feature analysis module with lightweight reinforcement learning (RL) for dynamic kernel selection, leveraging multi-scale feature pyramids and a Q-learning model optimized for GPU parallel execution; (2) a dynamic resource allocation strategy that partitions GPU resources by kernel size, predicts register/shared memory requirements, and optimizes memory access patterns; and (3) a pipeline folding and cache-aware scheduling framework that integrates feature analysis, decision-making, and convolution into a single kernel while enabling priority-driven preemption for real-time performance. Extensive experiments on datasets like CIFAR-100, ImageNet-1K, and VOC2012 demonstrate that the framework achieves up to 39% higher inference throughput, 2.2% improved recognition accuracy, and 16% better GPU utilization compared to state-of-the-art methods, all while maintaining sub-30[Formula: see text]ms end-to-end latency. The proposed approach bridges the gap between adaptive kernel intelligence and GPU efficiency, offering a robust solution for real-time visual pattern recognition tasks.