Small Object Detection in Fundus Images

In the field of pattern recognition, the accuracy of detection of small objects is significantly lower in comparison with the accuracy of common object detection. This contribution presents the evaluation of various machine learning techniques focused on improving the detection of small objects. We utilize these methods in medical imaging to diagnose eye diseases (diabetic retinopathy). We trained 5 modifications of the YOLOv8 model by applying different data preprocessing and improving their performance by integrating the SAHI slicing window algorithm. The use of SAHI led to a significant improvement in detection, reaching up to 16.37 % relative increase in mAP by IoU=0.5 and up to 30.93 % relative mAP increase by IoU = 0.1. This performance shows the high efficiency of the SAHI method in small object detection performance.