Detection and Recognition for Multiple Flames Using Deep Learning

• Main Author: Xin, Chen (Author)
• Other Authors: Yan, Wei Qi (Contributor), Nguyen, Minh (Contributor)
• Format: Others
• Published: Auckland University of Technology, 2018-11-01T23:01:22Z.
• Subjects: Flame classification; Flame recognition; Deep learning; Convolutional Neural Networks (CNN); Tensorflow; Single Shot MultiBox Detector (SSD); You Only Look Once (YOLO); R-FCN; Thesis
• Online Access: Get fulltext

 Object recognition is one of the critical fields of intelligent surveillance which can be applied to tackle multiple security issues. The recognition of a flame is an extended research direction which has significant practical value for multiple applications. The main contribution of this thesis, with the starting point of recognizing flames, is to introduce how Convolution Neural Networks (CNNs) deal with surveillance images. To achieve this, we constructed different neural network models, e.g., Single Shot MultiBox Detector (SSD) and You Only Look Once (YOLO). We enhanced the capacity of transformation of CNNs in part of convolution and RoI pooling. In this thesis, we mainly present assorted methods in deep leaning to achieve detection and identification of various flames. First, we introduce the fundamental theories of object detection and recognition using deep learning. After that, we detail the algorithms of pattern classification and compare those algorithms for image feature extraction. Based on the theory of neural networks, we also depict these learning methods and analyze the accuracy of experimental results. The focus of this thesis is on flame recognition using deep learning. For achieving this goal, a significant amount of measured data was used as the input data to ensure the validation of our experiments. We employ TensorFlow as the development environment with an NVIDIA GPU to train the image dataset and construct a detection model for the test dataset. Ultimately, experimental results show that the verification accuracy of our proposed algorithm is up to 98.6%.