基于光学镜头的表面缺陷实时检测算法

打开文本图片集

中图分类号：TP391.4 文献标识码：A 文章编号：2096-4706（2026）03-0127-06

Real-time Detection Algorithm for Surface Defects Based on Optical Lenses

HAN Junru， GUO Jianfeng （Collegeof Science and TechnologyNingbo University，Ningbo 3153o0，China）

Abstract： YOLO series models have become the benchmark in the real-time detection domain due to their balanced eficiencyandacuracyyet theirperformanceisrestricted bytheinerentfawsofNMSpost-procesing，incudingadiioal computationaloverheadandtheeroneous eliminationofvalidinformation.Althoughend-to-end Transformer frameworkscan breakfreefromtherelianceonNMS，theirexcessvelyhighcomputationalcostlimitspracticalapplication.Toadrsthis，the studyproposes adual-dimensional inovativearchitecture，whichconstructsadedicatedTransformernetwork forreal-timeendto-end detection and develops a hybrid Mamba-Transformer backbone network named Mamba Vision，significantly enhancing the multi-scalefeaturerepresentationcapability.Aphasedoptimizationstrategyisadopted，withparalelacelerationacieved bydecoupling intra-scaleand cross-scale feature processng for speed improvement and anuncertainty-minimization query initializationstrategyintroducedforaccuracyenancementtoboost theauracyoftargetlocalization.OntheCOCObenchmark， the proposed model achievesa mean Average Precision （mAP） of 53.05% with an inference speed exceeding 100FPS on a T4 GPU.Compared with mainstream YOLO variants，it improves detection accuracy by 4.2%～6.7% ，achieves breakthroughs in both speed and accuracy and thus provides a novel paradigm for industrial real-time detection systems.

Keywords：optical lens; real-time Object Detection; end-to-end

0 引言

实时目标检测是一个重要的研究领域，其应用范围广泛，涵盖目标跟踪[1]、视频监控[2]和自动驾驶[3]等多个方向。（剩余11705字）