Amid the increasing urgency of global waste management challenges, waste sorting has garnered significant attention. In this context, the integration of deep learning algorithms into waste categorization has rapidly expanded, driven by their significant improvements in accuracy and operational efficiency.However, contemporary deep learning frameworks require considerable computational resources, which poses a substantial barrier to their integration into small terminal devices. To overcome this obstacle, we present YOLOv8 Part Waste Sorting Correct Guidance (YOLOv8-PSCG), a refined model for domestic waste detection, meticulously designed as follows: starting with the implementation of ghost convolution to reduce computational complexity, thereby enhancing inference speed by transforming the traditional convolutional paradigm. We then devised the Faster Implementation of the CSP Bottleneck with 2 convolutions-Part Specific Classification (C2F-PSC) structure by enhancing the C2F framework through the incorporation of spatial and channel reconstruction convolution (SCConv) along with Rep-VGG, which reduces the overall complexity of the model and eliminates redundant functionalities. In the final stage of the design, the Asymptotic Feature Pyramid Network (AFPN) module is integrated into the network to facilitate seamless information exchange between different layers, mitigate information loss across multiple layers, and enhance understanding of crucial and comprehensive details. Empirical analysis performed on the domestic waste dataset, provided by the "Huawei Cloud Artificial Intelligence Competition - Waste Sorting Challenge Cup", confirms that our model has achieved a notable detection accuracy of 86.3% (mAP@50 : 95), significantly surpassing other state-of-the-art algorithms. Furthermore, the YOLOv8-PSCG model reduces the number of parameters by 5.3M and Floating Point Operations (<InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11760_2025_3827_Article_IEq1.gif" Format="GIF" Height="14" Rendition="HTML" Resolution="72" Type="Linedraw" Width="52" /> </InlineMediaObject> <EquationSource Format="TEX">$\textrm{FLOPs}$</EquationSource> <EquationSource Format="MATHML"><math> <mtext>FLOPs</mtext> </math></EquationSource> </InlineEquation>) by 0.8G, while improving overall performance by 1.2% (mAP@50 : 95). Consequently, YOLOv8-PSCG achieves a harmonious balance between a streamlined architecture and improved accuracy, making it highly suitable for deployment on small terminal devices. These improvements improve the model’s practicality and open new possibilities for applying deep learning in resource-constrained environments.

Improved lightweight deep learning architecture for household waste detection

期刊论文
发表日期：

2025年1月28日

查看全文

Qin Dong,
Jun Xue,
Shuting Chen,
Jinxin Zhu

摘要

Amid the increasing urgency of global waste management challenges, waste sorting has garnered significant attention. In this context, the integration of deep learning algorithms into waste categorization has rapidly expanded, driven by their significant improvements in accuracy and operational efficiency.However, contemporary deep learning frameworks require considerable computational resources, which poses a substantial barrier to their integration into small terminal devices. To overcome this obstacle, we present YOLOv8 Part Waste Sorting Correct Guidance (YOLOv8-PSCG), a refined model for domestic waste detection, meticulously designed as follows: starting with the implementation of ghost convolution to reduce computational complexity, thereby enhancing inference speed by transforming the traditional convolutional paradigm. We then devised the Faster Implementation of the CSP Bottleneck with 2 convolutions-Part Specific Classification (C2F-PSC) structure by enhancing the C2F framework through the incorporation of spatial and channel reconstruction convolution (SCConv) along with Rep-VGG, which reduces the overall complexity of the model and eliminates redundant functionalities. In the final stage of the design, the Asymptotic Feature Pyramid Network (AFPN) module is integrated into the network to facilitate seamless information exchange between different layers, mitigate information loss across multiple layers, and enhance understanding of crucial and comprehensive details. Empirical analysis performed on the domestic waste dataset, provided by the "Huawei Cloud Artificial Intelligence Competition - Waste Sorting Challenge Cup", confirms that our model has achieved a notable detection accuracy of 86.3% (mAP@50 : 95), significantly surpassing other state-of-the-art algorithms. Furthermore, the YOLOv8-PSCG model reduces the number of parameters by 5.3M and Floating Point Operations ( $\textrm{FLOPs}$ $FLOPs$ ) by 0.8G, while improving overall performance by 1.2% (mAP@50 : 95). Consequently, YOLOv8-PSCG achieves a harmonious balance between a streamlined architecture and improved accuracy, making it highly suitable for deployment on small terminal devices. These improvements improve the model’s practicality and open new possibilities for applying deep learning in resource-constrained environments.