Development of an algorithm for formulating recommendations for the selection of object detection models based on data meta-features and an experimental knowledge base

Object detection in video surveillance systems, medical diagnostics, autonomous vehicles, and other technical systems incorporating computer vision components is closely tied to selecting the optimal model for a specific dataset. This task remains labor-intensive and requires substantial computational and time resources. In this work, the authors propose a method for automated generation of recommendations for selecting an object detection model based on the analysis of image meta-features and a knowledge base of results from previous studies. The proposed method involves forming meta-feature vectors from a dataset (images and annotations) and searching for similar datasets within a collected knowledge base using the CatBoostClassifier algorithm (F1-score = 0.82). After identifying analogous datasets, object detection models are ranked by accuracy (mAP50) and performance (FPS) based on previously accumulated experimental data, enabling the recommendation of the most suitable architectures. The method is implemented in the open-source recommendation framework Object Detection Recommendation System (ODRS). Experimental evaluation confirmed that applying the proposed algorithm reduces model selection time by 40 % compared to traditional approaches that require manual testing. Validation on 12 test datasets demonstrated high efficiency of the method in tasks with limited computational resources and diverse image characteristics. Due to its integration into ODRS, the method is accessible to specialists without deep expertise in machine learning or computer vision. It is particularly useful in scenarios with constrained time and computational resources, where a quick, well-founded initial approximation for model selection is needed. Future work includes expanding the knowledge base and adapting the method for semantic segmentation and object tracking tasks.