An integrated approach to object detector compression and evaluation using the YOLO example

The article presents a software module for complex optimization of object detection models using the YOLO11 architecture as an example. The developed system implements a full cycle of working with the model — training, pruning, fine-tuning, quantization and subsequent comparative analysis. During the experiments, two modifications of YOLO11m and YOLO11n were evaluated using the CompDetect and Safety Helmet Detection datasets, which differ in complexity and image structure. The results showed that the use of the proposed module reduces computational costs while maintaining a high level of detection accuracy. The presented approach can be used as a universal tool for optimizing and implementing neural network detectors in real-time systems.