GENERALIZED APPROACH FOR OBJECT SELECTIVE SEARCH IN IMAGES

Abstract

The work is devoted to the process of research and development of an own approach for object recognition in images in cases of selective search. In the modern world in the field of computer vision and image processing, object recognition is one of the most important areas of research. The use of neural networks, such as YOLO (You Only Look Once) and R-CNN (Region-based Convolutional Neural Network), has proven to be very effective in solving this problem. These algorithms are able to find objects in images and return bounding boxes that accurately describe those objects. However, in some cases when we work with interactive programs, for example, selecting an object by clicking or touching a recognized area, there is a problem of selecting the correct object and its bounding box. This can affect the accuracy of the selected object in the context of a selective search. It is necessary to find such a search area that will allow us to properly identify the selected object, especially in the case of crossing bounding boxes. An effective approach to the size of the search area and the visualization of the research process can improve the accuracy and speed of object selection, providing a more convenient and efficient object search in images. We offer a solution to the problem of crossing the bounding frames that arises in the operation of YOLO and R-CNN type neural networks by developing a method for estimating the optimal size of the search area, which will allow finding the appropriate object and its bounding frame, and offer a generalized approach to the visualization of the research process, which will allow to visually represent the overlap of the bounding frames and will facilitate the selection of the optimal object. To confirm the effectiveness of the proposed method, we conduct experiments on the appropriate data set and compare them. The results of such research can have a significant practical impact on the development of object recognition systems and improving their functionality as a whole. Future research may focus on expanding the dataset for selective search cases, including different bounding box overlap scenarios and objects with different shapes and sizes.

Keywords: Drones, selective object search, bounding boxes, neural networks, YOLO, R-CNN, recognition algorithm.