Improving object detection performance by lightweight approaches

Abstract

Object Detection has been a significant topic in computer vision. As the continuous development of Deep Learning, many advanced academic and industrial outcomes are established on localising and classifying the target objects, such as instance segmentation, video tracking and robotic vision. As the core concept of Deep Learning, Deep Neural Networks (DNNs) and associated training are highly integrated with task-driven modelling, having great effects on accurate detection. The main focus of improving detection performance is proposing DNNs with extra layers and novel topological connections to extract the desired features from input data. However, training these models can be a computational expensive and laborious progress as the complicated model architecture and enormous parameters. Besides, dataset is another reason causing this issue and low detection accuracy, because of insufficient data samples or difficult instances. To address these training difficulties, this thesis presents two different approaches to improve the detection performance in the relatively light-weight way. As the intrinsic feature of data-driven in deep learning, the first approach is "slot-based image augmentation" to enrich the dataset with extra foreground and background combinations. Instead of the commonly used image flipping method, the proposed system achieved similar mAP improvement with less extra images which decrease training time. This proposed augmentation system has extra flexibility adapting to various scenarios and the performance-driven analysis provides an alternative aspect of conducting image augmentation. The "StomaRCNN" is the second approach which is based on a realistic application task to automatically detect, segment and measure the stomata in plant microscope images. The key innovation of StomaRCNN is reorganising DNN pipeline to utilise the detailed features in high-resolution microscope images without damaging the image qualities. Despite the limited related works, StomaRCNN achieved human-level measurement accuracy for open stoma instances and demonstrates a large potential of applying Deep Learning approaches to automatically solve instance measuring problems in plant science. Those presented works propose alternative ways of improving object detection performance and highlight the importance of rethinking object detection in the aspects of data-driven and step-wise architectural design.

More information

Identifiers

ORCID for Yingwei Zhou: orcid.org/0000-0001-6131-3156

Catalogue record

Export record