UNIFIED REAL TIME OBJECT DETECTION USING CONVOLUTIONAL NEURAL NETWORK

Abstract

One of the most important parts of image processing system is object detection. The main objective of object detection is to find object location. There are two techniques developed to extract features of the object in object detection system: knowledge-based and learning based techniques. The obtained results of recent works showed that the traditional YOLO and R-CNN are the most suitable algorithms for real-time and low-error rate object detections. YOLO is considered as a significant algorithm for computer vision applications due to its high speed with a good accuracy for real time object detection. However, YOLO has several limitations such as localization errors and poor small object detection performance and in variations in object orientation. In this thesis, two algorithms are proposed, designed, and evaluated to address and overcome the drawbacks and limitations of the traditional YOLO. The first proposed algorithm enhances the performance of YOLO-V2 by modifying the backbone network, anchor box prediction method, and loss function to efficiently detect oil tanker trucks. The second proposed algorithm enhances the performance of YOLO-V3 by designing and building a CNN to solve the problem of large number of YOLO-V3 parameters, using densely connected modules to enhance the interlayer connection of CNNs and further strengthen the connection between dense neural network blocks, and finally improving the YOLO-V3 multiple-scale detection by expanding the three-scale to four-scale detection to increase the accuracy of detecting small objects like drones. It can be concluded from the obtained results of the evaluation tests that the main mAP of the first designed algorithm (OYOLOv2_FTD) improves the detection performance to reach 93.0%. The developed algorithm (OYOLOv2_FTD) is superior over the original (YOLOV2) by 4% in detecting oil tanker trucks. The proposed algorithm is more suitable for real-time detecting and tracking oil tanker trucks. Moreover, evaluation results of the second designed algorithm provide good robustness to drone detection. As well as the accuracy and average precision reached 95.60 and 96%, respectively. This improves the feasibility and efficiency of YOLO-V3 for drone objects in the image’s detection.