Development of a Detector for Stamps on Images

In this paper, various methods and algorithms are investigated to develop an effective stamp detector that can automatically find and highlight various stamps in images with high speed and accuracy. The article uses various approaches, including machine learning and computer vision. A proprietary machine learning model (pre-trained) based on new data was created. In addition, the entire process of collecting and preparing data for training models is described, including image augmentation methods to improve the generalization ability of models. Various metrics for assessing the quality of detection, such as accuracy, recall and mAP50, are considered, as well as how they can be used to select the optimal model. A comparison is also made of various approaches to solving the problem of detecting stamps on images, including their advantages and disadvantages. An experimental comparison of various YOLOv8 models was carried out on a validation data set and conclusions were drawn about the effectiveness of each in the context of the task.