<p>To address challenges in dense product placement, incomplete feature capture in automatic vending cabinets, and image distortion caused by fisheye cameras, a novel retail cabinet product recognition algorithm based on YOLOv8 is proposed. The original YOLOv8 model is enhanced with the feature fusion attention network (FFA-Net) to improve detection performance for blurred images. Additionally, optimized CReToNeXt blocks are designed and integrated to replace the original C2f blocks in the head section, and SlideLoss is introduced to optimize the training process. Experiments were conducted using a public retail cabinet dataset, including comparative and ablation studies to evaluate the effects of each improvement. The results demonstrate that the enhanced YOLOv8 model achieves a recognition precision improvement from <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11760_2025_4180_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="43" /> </InlineMediaObject> <EquationSource Format="TEX">$96.3\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>96.3</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> to <InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11760_2025_4180_Article_IEq2.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="43" /> </InlineMediaObject> <EquationSource Format="TEX">$98.6\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>98.6</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> and an mAP of <InlineEquation ID="IEq3"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11760_2025_4180_Article_IEq3.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="43" /> </InlineMediaObject> <EquationSource Format="TEX">$99.0\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>99.0</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation>. These findings confirm that the proposed algorithm is highly effective for high-precision detection and recognition in unmanned vending cabinet scenarios. The research code is publicly available on GitHub at <a href="https://github.com/112345434/shangpinshibie">https://github.com/112345434/shangpinshibie</a>.</p>

Enhanced YOLOv8 for high-precision retail cabinet product recognition

期刊论文
发表日期：

2025年5月13日

查看全文

Yachao Si,
Jiajie Gao,
Xingxuan Zhao,
Xiaojun Xu

摘要

To address challenges in dense product placement, incomplete feature capture in automatic vending cabinets, and image distortion caused by fisheye cameras, a novel retail cabinet product recognition algorithm based on YOLOv8 is proposed. The original YOLOv8 model is enhanced with the feature fusion attention network (FFA-Net) to improve detection performance for blurred images. Additionally, optimized CReToNeXt blocks are designed and integrated to replace the original C2f blocks in the head section, and SlideLoss is introduced to optimize the training process. Experiments were conducted using a public retail cabinet dataset, including comparative and ablation studies to evaluate the effects of each improvement. The results demonstrate that the enhanced YOLOv8 model achieves a recognition precision improvement from $96.3\%$ $96.3 %$ to $98.6\%$ $98.6 %$ and an mAP of $99.0\%$ $99.0 %$ . These findings confirm that the proposed algorithm is highly effective for high-precision detection and recognition in unmanned vending cabinet scenarios. The research code is publicly available on GitHub at https://github.com/112345434/shangpinshibie.