Overcoming resolution constraints in automated colony counting via a high-performance deep learning framework using SAHI.

Lightweight models perform poorly in bacterial colony counting when high-resolution Petri dish images are downscaled to 640 × 640 pixels. This study addresses this issue using a tiled training and SAHI-based tiled inference pipeline. Full-resolution images are divided into overlapping 640 × 640 tiles (20% overlap), preserving native resolution and reducing colony density per tile. A public 24-class bacterial colony dataset was used to evaluate the proposed approach. The tiled strategy substantially increased the number of training samples while preserving local colony details. Three nano-scale models (YOLOv5n, YOLOv8n, YOLOv11n) were evaluated. The conventional resize method yielded only 44.9-66.3% mAP@0.5, whereas the tiled approach achieved 95.4-96.9% mAP@0.5 and 55.4-58.4% mAP@0.5:0.95. YOLOv11n with tiling achieved the highest mAP@0.5 while using only 2.6 million parameters. It reached 96.9% mAP@0.5 and 58.2% mAP@0.5:0.95, and the validation-set confusion matrix showed class-wise correct prediction rates above 95% in 23 of 24 classes. Whole-plate inference required less than 320 ms on a laptop RTX 3050 GPU. For small, densely packed objects such as bacterial colonies, tiled inference with moderate overlap dramatically outperforms both image resizing and recent architectural advances in YOLO. The proposed lightweight pipeline is suitable for routine laboratory deployment.