MKNet: Adaptive Kernel Mixing Depthwise Convolution for Lightweight Object Detection

doi:10.1109/jiot.2026.3725353

MKNet: Adaptive Kernel Mixing Depthwise Convolution for Lightweight Object Detection

2026-08-20

0

PRE

AI

DOI：10.1109/jiot.2026.3725353

Original

Original request for help

Save

Abstract

En 中文

Lightweight object detection has become increasingly important for deployment in resource-constrained environments such as embedded systems and edge devices, as it enables instantaneous decision-making, reduces data transmission latency, and improves the reliability of real-world operations. Although recent lightweight detectors reduce model complexity through depthwise separable convolutions, channel grouping, and structural simplification, these strategies often weaken cross-channel interaction and contextual feature representation, leading to performance degradation in complex scenes. To address this limitation, this paper proposes a novel lightweight object detection architecture, MKNet, that enhances spatial diversity and feature interaction while preserving parameter efficiency. Specifically, we introduce an Adaptive Kernel Mixing Depthwise (AKMixDW) module that enriches spatial representation through multi-kernel depthwise filtering, enabling improved contextual modeling without significant parameter growth. Furthermore, a Partial Standard Convolution mechanism is integrated to strengthen cross-channel interaction under compact structural constraints, mitigating the representational limitations commonly observed in lightweight designs. Extensive experiments conducted on benchmark datasets demonstrate that the proposed method achieves a favorable accuracy–efficiency trade-off compared to existing lightweight baselines. MKNet achieves 46.9% mAP@0.5 and 31.6% mAP@0.5:0.95 with 1.99M parameters and 7.8 GFLOPs, while attaining 43.66 FPS on the NVIDIA Jetson Orin Nano. Ablation studies further validate the effectiveness of each component in improving detection robustness while maintaining fast inference capability. The proposed framework provides an effective means of balancing detection performance and parameter efficiency in practical deployment scenarios.

Keywords:

Deep Learning

Lightweight object detection

YOLO

MobileNetV3

MKNet: Adaptive Kernel Mixing Depthwise Convolution for Lightweight Object Detection

Abstract

Journal

Organization

Cited Papers