چکیده مقاله
Object detection is one of the main tasks in computer vision This is an essential and challenging problem in computer vision Small object detection plays a fundamental role in many computer vision tasks In recent years, the successes achieved in deep learning techniques have led to the discovery of new ways to solve the problem of object detection has shrunk and pushed it into prominent research Extensive studies in the field of object detection have been studied and used in universities and their applications in the real world, such as robot vision, autonomous driving, intelligent transportation, drone vision analysis, military reconnaissance, and surveillance The experimental results show that the accuracy of small object detection with methods based on region proposal has performed best, combining several methods such as classification and segmentation in detection Small objects can facilitate performance Object detection based on deep learning and segmentation has achieved unprecedented progress In this research, we have used Mask R CNN to improve small object detection Here, we used Distance IOU, which in addition to calculating the overlap of two boxes, tries to match the distance between two boxes instead of the original loss function, and also we used Cluster NMS instead of the original maximum suppression which led to an increase of 0 71% in average precision in bounding box head and 0 26% in average precision in segmentation head for small objects compared to Mask R CNN with original Non Maximum Suppression and Intersection Over Union loss function We have made this improvement using the Detectron2 platform
کلیدواژهها
نویسندگان
Mahsa Nobakht
Department of Computer Science, Najafabad Branch, IslamicAzad University, Najafabad, Iran
Mehdi Amoon
Department of Eleectrical Engineering, Najafabad Branch,Islamic Azad University, Najafabad, Iran
شیوه ارجاع
Nobakht, Mahsa and Amoon, Mehdi,1401,An improved method based on convolutional network to increase of accuracy of small object detection in complex images,Second International Conference on Computer Engineering and Science,Najafabad
ارائهشده در
مجموعه مقالات دومین کنفرانس بین المللی مهندسی و علوم کامپیوتر29 آذر 1401 · نجف آباد