Small object recognition using machine learning techniques

• Main Author: Loy, Hui Juan
• Other Authors: Huang Guangbin
• Format: Final Year Project
• Language: English
• Published: 2017
• Subjects: DRNTU::Engineering::Electrical and electronic engineering::Computer hardware, software and systems
• Online Access: http://hdl.handle.net/10356/71119
• Institution: Nanyang Technological University

In recent years, there has been significant interest in deep machine learning, due to its flexibility of application for different fields. In particular, autonomous vehicles apply deep machine learning algorithm for object detection as an integral component of its functionality. To ensure that the autonomous vehicles can react to road related events, the deep machine learning algorithm must be able to detect small objects. However, insufficient visual information within images causes small objects to be classified as noises during the detection process, resulting in the failure in detection. Therefore, the objective of this project is to detect small objects using deep machine learning techniques adapted from pre-existing models. The first task of this project considered different models of varying accuracy and computation time, specifically You Only Look Once (YOLO) and Single Shot Detector (SSD), to a set criteria derived from pre-existing standards. SSD was eventually chosen because of its higher accuracy and the relative ease of reducing computation time. The second part of the project focused on training the model for small object detection. The first task was to prepare a dataset for training. From 2 existing training datasets (VOC2007, VOC2012), a new dataset of random images, which consists of small objects from 20 different classes was extracted, using a predefined set of annotation files that located small objects within the 2 datasets. However, the outcome was not favourable; due to the size limitations of SSD, the dataset was insufficient to train the model to even achieve already established accuracy. The Super Resolution (SR) method from Fast super resolution convolutional neural network(FSRCNN) was then introduced to prepare the images before being processing by SSD. Using a new dataset, this additional method to pass the images through SR, then SSD, achieved a higher accuracy of detecting small objects than the previous outcome.