Systematic Evaluation of Object Detection Model Modifications for Small-Dataset Apple Surface Defect Detection

Abstract

Automatic detection of apple surface defects plays a crucial role in precision agriculture. However, developing an optimal detection model for this task remains challenging because defect sizes vary widely, ranging from small punctate lesions to large discolored patches, while the scarcity of annotated training data further limits the applicability of general-purpose detectors. To identify the most effective module combination for small-dataset defect detection, we systematically evaluate more than 50 configurations by combining three architectural modifications, namely the P2 head, BiFPN neck, and CBAM, with three loss-function modifications, including positive-class weighting, EIoU, and Focal Loss, using YOLOv8-s as the base model. Experiments are conducted on a laboratory-collected dataset containing 30 test images with 80 ground-truth instances, and performance is evaluated using mAP at $\text{IoU} = \text{0.5}$. The results show that CBAM combined with positive-class weighting achieved the most practically balanced result in terms of mAP (94.71%), Recall@0.5 (86.2%), and FP count (112), outperforming the YOLOv8-s+P2 baseline, which achieves 93.51% mAP and 77.5% Recall@0.5. Furthermore, the combination of BiFPN, CBAM, and positive-class weighting degraded mAP to 87.48%, indicating harmful interference among the randomly initialized modules under the amplified loss. These findings provide practical insights into designing effective detection models for small-dataset defect detection scenarios.