Real-time anomaly detection in industrial production is vital for ensuring both safety and operational efficiency. However, conventional object detectors often struggle in such environments due to the extremely small scale of targets and pervasive background interference. In particular, tiny aluminum cans-occupying only a small fraction of the frame-are easily missed when motion blur and standard downsampling erode their fine-grained details, while uneven lighting and mechanical vibrations further diminish detection robustness. To overcome these limitations, we introduce YOLOv8-CSPD, a tailored enhancement of YOLOv8 designed for reliable small-object detection in challenging industrial settings. Our model embeds the Convolutional Block Attention Module (CBAM) to deliver adaptive channel-and-spatial attention, selectively amplifying relevant textures and edges while suppressing distracting noise. Concurrently, we replace conventional strided convolutions with a Spatial Pyramid Depthwise Separable Convolution (SPDConv) - a lossless space-to-depth transformation that preserves all critical texture information across scales. Evaluations on a bespoke industrial aluminum-can dataset reveal that YOLOv8-CSPD achieves markedly superior precision and recall compared to both the original YOLOv8x and other leading detectors such as YOLOv9 and RE-DETR, all while retaining real-time inference speeds. These outcomes demonstrate that the synergistic fusion of collaborative attention and lossless downsampling substantially bolsters small-object detection performance, rendering YOLOv8-CSPD a practical solution for high-throughput, real-time industrial inspection.

错误:搜索内容不能为空,请输入英文关键词
错误:关键词超出字数限制,请精简
高级检索

YOLOv8-CSPD: Integrating Collaborative Attention and Lossless Downsampling for Real-Time Industrial Aluminum Can Detection

  • Wenjie Qian,
  • Liyun Liu,
  • Luokun He,
  • Zhixuan Li,
  • Xiao Wang

摘要

Real-time anomaly detection in industrial production is vital for ensuring both safety and operational efficiency. However, conventional object detectors often struggle in such environments due to the extremely small scale of targets and pervasive background interference. In particular, tiny aluminum cans-occupying only a small fraction of the frame-are easily missed when motion blur and standard downsampling erode their fine-grained details, while uneven lighting and mechanical vibrations further diminish detection robustness. To overcome these limitations, we introduce YOLOv8-CSPD, a tailored enhancement of YOLOv8 designed for reliable small-object detection in challenging industrial settings. Our model embeds the Convolutional Block Attention Module (CBAM) to deliver adaptive channel-and-spatial attention, selectively amplifying relevant textures and edges while suppressing distracting noise. Concurrently, we replace conventional strided convolutions with a Spatial Pyramid Depthwise Separable Convolution (SPDConv) - a lossless space-to-depth transformation that preserves all critical texture information across scales. Evaluations on a bespoke industrial aluminum-can dataset reveal that YOLOv8-CSPD achieves markedly superior precision and recall compared to both the original YOLOv8x and other leading detectors such as YOLOv9 and RE-DETR, all while retaining real-time inference speeds. These outcomes demonstrate that the synergistic fusion of collaborative attention and lossless downsampling substantially bolsters small-object detection performance, rendering YOLOv8-CSPD a practical solution for high-throughput, real-time industrial inspection.