Impact of image segmentation on video frame classification performance with tiny objects

This paper introduces a novel algorithm to classify video frames with tiny informative objects that represent less than 1.5% of the scene. Utilizing advanced image segmentation and data preprocessing techniques, the algorithm improves neural networks’ generalization and classification performance for images dominated by minor significant objects. The efficacy of this method is validated against two existing algorithms through comprehensive experiments, demonstrating improved accuracy in detecting small objects. This research has broad implications, extending from sports analytics to surveillance, microscopic medical imaging, and satellite imagery, by offering a robust solution for classifying minimal object features in various domains.