基于并联双注意力的轻量级小样本矿石粒度检测

打开文本图片集

关键词：计算机视觉；小样本目标检测；轻量化；矿石图像；实时检测中图分类号：TP394.1；TH691.9 文献标识码：Adoi：10.37188/OPE.20263402.0309 CSTR：32169.14.OPE.20263402.0309

A lightweight few-shot ore detector with parallel dual attention

SUN Guodong 1，2* ，LIU Mingxuan1，LI Shicheng4，WU Bo³

（1.College ofMechanical Engineering，Hubei University of Technology，Wuhan 43Oo68，China;

2.Hubei Key Laboratory of Modern Manufacturing Quality Engineering，Hubei University of Technology，Wuhan 430068，China;

3. Shanghai Aduanced Research Institute， Chinese Academy ofSciences， Shanghai 2Ol21O，China;

4.Detroit Green Technology Institute，Hubei University of Technology，Wuhan 43O068，China） * Corresponding author， E -mail： sgdeagle@163. com; wubo@sari. ac. cn

Abstract： To address the high computational complexity，limited feature robustness，and constrained classifier performance of conventional object detection methods in ore particle size detection，a few-shot object detection approach was proposed to reduce annotation cost and improve generalization under data-scarce conditions.The proposed method was built upon the CenterNet2 framework and employed a lightweight VoVNet as the backbone to ensure detection eficiency. A parallel dual-attention feature fusion module was designed as the core component. Specifically，a channel cross-attention module was introduced to re calibrate channel-wise feature responses，while a spatial group-atention module emphasized discriminative target regions. The coordinated operation of the two modules enhanced the fusion of task-relevant features and provided effective guidance for query image detection in few-shot scenarios. Experimental results on an ore dataset show that the proposed model achieved an average precision （AP） of 55.2% ，with AP50 and AP75 reaching 78.5% and 66.9% ，respectively. The inference speed reached 57 frames per second （FPS），while the attention module required only 16.1M parameters， indicating a favorable trade-off between accuracy and eficiency. Experimental results demonstrate that the proposed method efectively enhances the perception performance of few-shot ore particle size detection. Moreover，it possesses high potential for edge deployment，providing a reliable technical solution for real-time detection challnges in smart mines under computation-constrained conditions.

Key words： computer vision; few-shot object detection; lightweight; ore images;real-time

1引言

随着科技的进步和工业化进程的加速，对矿石资源的需求日益增长，高效、准确地识别矿石粒度大小成为提高开采效率、优化资源利用的关键环节。（剩余18327字）