YOLOv5 + Flask

Spacer Defect Detection

A computer vision system that detects 23 classes of defects in spacers, served through a Flask web app with bounding-box results and CSV export.

0.56

mAP@0.5

0.431

mAP@0.5:0.95

0.92

Best class mAP@0.5

The problem

Manual inspection of spacers is slow and inconsistent. The dataset was small — 1,015 labeled images across 23 defect classes — so the model had to generalize from little data.

Approach

• Transfer learning from the YOLOv5s checkpoint.
• Heavy data augmentation, expanding the training set to over 150,000 images.
• Flask app for uploading an image, viewing detections with bounding boxes and downloading results as CSV for quality-control systems.

Stack

• YOLOv5
• PyTorch
• Flask
• Computer Vision