Deep learning object detector using a combination of Convolutional Neural Network (CNN) architecture (MiniVGGNet) and classic object detection algorithm

• Main Authors: Ismail, Asmida, Ahmad, Siti Anom, Che Soh, Azura, Hassan, Mohd Khair, Harith, Hazreen Haizi
• Format: Article
• Language: English
• Published: Universiti Putra Malaysia Press 2020
• Online Access: http://psasir.upm.edu.my/id/eprint/81052/1/CNN.pdf; http://psasir.upm.edu.my/id/eprint/81052/; http://www.pertanika.upm.edu.my/pjst/browse/special-issue?article=JST(S)-0566-2020

The object detection system is a computer technology related to image processing and computer vision that detects instances of semantic objects of a certain class in digital images and videos. The system consists of two main processes, which are classification and detection. Once an object instance has been classified and detected, it is possible to obtain further information, including recognizes the specific instance, track the object over an image sequence and extract further information about the object and the scene. This paper presented an analysis performance of deep learning object detector by combining a deep learning Convolutional Neural Network (CNN) for object classification and applies classic object detection algorithms to devise our own deep learning object detector. MiniVGGNet is an architecture network used to train an object classification, and the data used for this purpose was collected from specific indoor environment building. For object detection, sliding windows and image pyramids were used to localize and detect objects at different locations, and non-maxima suppression (NMS) was used to obtain the final bounding box to localize the object location. Based on the experiment result, the percentage of classification accuracy of the network is 80% to 90% and the time for the system to detect the object is less than 15sec/frame. Experimental results show that there are reasonable and efficient to combine classic object detection method with a deep learning classification approach. The performance of this method can work in some specific use cases and effectively solving the problem of the inaccurate classification and detection of typical features.