<p>Trademarks are important identifiers for goods and services. They play an increasingly important role in daily life and production. However, with the continuous development of commercial society and the increasing number of goods and services, trademark detection and classification technology has become particularly critical, and it has large application prospects in the fields of commodity image search and trademark infringement. Based on the You Only Look Once (YOLO)-v5 framework, this paper proposes the YOLO-BC network to better detect trademark images. First, the network adopts a new bi-directional feature pyramid network & coordinate attention(BiFC) module to replace the original enhanced feature extraction network (PANet) model to improve the feature extraction ability of the network at different scales and capture key information. Second, a convolutional block attention module (CBAM) is introduced into the C3 module, and the influence of the C3-CBAM module embedded in the backbone and neck of the algorithm network on network performance is explored through experiments. Subsequently, the distance intersection over union non-maximum suppression (DIOU-NMS) is used to replace the NMS in the original network as the bounding box non-maximum suppression method. Finally, we replace generalized IOU (GIOU) with <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11042_2025_20870_Article_IEq1.gif" Format="GIF" Height="10" Rendition="HTML" Resolution="72" Type="Linedraw" Width="14" /> </InlineMediaObject> <EquationSource Format="TEX">$ \alpha $</EquationSource> <EquationSource Format="MATHML"><math> <mi>α</mi> </math></EquationSource> </InlineEquation>-IOU as the bounding box regression loss function and use the appropriate parameter <InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11042_2025_20870_Article_IEq1.gif" Format="GIF" Height="10" Rendition="HTML" Resolution="72" Type="Linedraw" Width="14" /> </InlineMediaObject> <EquationSource Format="TEX">$ \alpha $</EquationSource> <EquationSource Format="MATHML"><math> <mi>α</mi> </math></EquationSource> </InlineEquation> to adaptively adjust the loss and gradient of the IOU target. Experimental results on public dataset and a self-constructed dataset show that the improved framework has a higher detection accuracy than traditional methods.</p>

Trademark detection and classification based on an improved YOLO

期刊论文
发表日期：

2025年4月24日

查看全文

Chunyuan Miao,
Xiuhui Wang

摘要

Trademarks are important identifiers for goods and services. They play an increasingly important role in daily life and production. However, with the continuous development of commercial society and the increasing number of goods and services, trademark detection and classification technology has become particularly critical, and it has large application prospects in the fields of commodity image search and trademark infringement. Based on the You Only Look Once (YOLO)-v5 framework, this paper proposes the YOLO-BC network to better detect trademark images. First, the network adopts a new bi-directional feature pyramid network & coordinate attention(BiFC) module to replace the original enhanced feature extraction network (PANet) model to improve the feature extraction ability of the network at different scales and capture key information. Second, a convolutional block attention module (CBAM) is introduced into the C3 module, and the influence of the C3-CBAM module embedded in the backbone and neck of the algorithm network on network performance is explored through experiments. Subsequently, the distance intersection over union non-maximum suppression (DIOU-NMS) is used to replace the NMS in the original network as the bounding box non-maximum suppression method. Finally, we replace generalized IOU (GIOU) with $ \alpha $ $α$ -IOU as the bounding box regression loss function and use the appropriate parameter $ \alpha $ $α$ to adaptively adjust the loss and gradient of the IOU target. Experimental results on public dataset and a self-constructed dataset show that the improved framework has a higher detection accuracy than traditional methods.