Enhancing Small Object Detection in the YOLOv8 model: A Comprehensive Analysis of the Optimized Model Head Adaptations

Strawberry fruits, with their petite dimensions, possess an innate ability to blend into the surrounding foliage seamlessly. The size differences between the fruit and the background are such that they pose significant recognition challenges event to deep neural network models (DNN). With this in mind, this work is focused on exploring the latest model in the YOLO family, specifically YOLOv8, to address the complexities of a dataset gathered via a high-throughput phenotyping platform. As such, this study proposed a comprehensive analysis of optimized model head adaptations to enhance the detection of small objects starting from the base YOLOv8 architecture. This optimization is pursued while retaining the effectiveness of the model in identifying larger fruits and flowers. This optimization is poised to yield positive effects and contribute to more accurate and robust detection results on the proposed dataset.