DBA-YOLOv5s wear detection method for lathe tool

Accurate visual detection of lathe-tool wear is challenging because wear regions can occupy small image areas, exhibit irregular morphologies and subtle inter-class differences, and be affected by surface reflections and background interference. To address these challenges, this study develops DBA-YOLOv5s, an improved YOLOv5s model that combines detection-scale refinement, bidirectional multi-scale feature fusion, and stage-specific attention. First, a 160×160 feature-map detection head is added to strengthen the representation of small wear regions. Second, additional cross-scale connections are introduced to shorten the information-transfer paths between shallow detail features and deep semantic features. Third, coordinate attention (CA) is embedded in selected backbone stages to preserve positional information, whereas efficient channel attention (ECA) is applied to the detection branches to refine channel responses with limited additional computation. The model is evaluated on a self-constructed dataset containing 3,510 images from nine lathe-tool wear and damage categories. Under the fixed 7:2:1 image-level split, DBA-YOLOv5s achieves a precision of 87.2%, a recall of 88.5%, and an [email protected] of 87.2%, compared with 80.3%, 81.2%, and 81.0% for YOLOv5s, corresponding to absolute improvements of 6.9, 7.3, and 6.2 percentage points, respectively. These results demonstrate the potential of the proposed task-oriented architecture under the acquisition conditions considered in this study.