Evaluation of bio-inspired pre-processing to improve object classification

Author: Ditish Maharjan

Maharjan, Ditish, 2021 Evaluation of bio-inspired pre-processing to improve object classification, Flinders University, College of Science and Engineering

Terms of Use: This electronic version is (or will be) made publicly available by Flinders University in accordance with its open access policy for student theses. Copyright in this thesis remains with the author. You may use this material for uses permitted under the Copyright Act 1968. If you are the owner of any included third party copyright material and/or you believe that any material has been made available without permission of the copyright owner please contact copyright@flinders.edu.au with the details.


High Dynamic Range (HDR) imaging helps to overcome the limited dynamic range of traditional imaging system, i.e. Low Dynamic Range (LDR) imaging. LDR imaging systems are capable of capturing less than three orders of dynamic information of a scene. Because of this limited dynamic range, LDR imaging systems cannot capture details in scenes with both dark and very bright regions in a single scene. This has been a limiting factor for implementing the LDR imaging system in Machine Vision (MV) applications such as agriculture, surveillance, autonomous driving.

The limitation of traditional imaging systems does not affect HDR imaging with its extended dynamic range. However, HDR imaging for its increased information content requires more storage, longer transfer time and computation power for its use of floating-point data to represent the dynamic range compared to 8-bit LDR images. Tone Mapping Operator (TMO) are used to dynamically compress the higher dynamic range of HDR images to 8-bit LDR images while still preserving some details. While there are multiple state-of-the-art TMOs available for such purpose, most of them have been designed with subjective metrics used for human consumption. There is a novel TMO designed with metrics to improve noise suppression, enhance image contrast and edge detection and reduce image flicker based on a biological inspiration from blowfly called bio-inspired TMO. The metrics used for bio-inspired TMO are focused on information content rather than artistic recreation for human consumption and hence more suited for MV applications.

This thesis is undertaken to evaluate the performance of bio-inspired TMO MV application of Object classification and Localisation on dynamic images. For evaluation, multiple datasets were captured in normal and low light condition on different camera setup using three HDR cameras: Monochrome, Colour and Low-Resolution Colour camera. The HDR images captured for each case were pre-processed using bio-inspired TMO to create PRC datasets. For comparison, the captured HDR images were also used to create LDR datasets by applying a gamma correction of 2.0 followed by histogram equalisation to create RAW datasets. Each of these datasets were annotated to create ground truth data. Faster R-CNN was used as the object detector to generate predictions on the datasets. These predictions were compared with the ground truth annotation data using evaluation metrics of PASCAL VOC.

Upon evaluating the results, it showed that on overall bio-inspired TMO helped to increased object detector’s performance. However, this increment was enjoyed only by datasets captured using Monochrome cameras. The gain for low resolution colour camera was marginal while in case of colour cameras the performance of object detector was found to decrease when using bio-inspired pre-processed datasets. Overall, applying bio-inspired TMO on datasets captured by Monochrome cameras showed the maximum improvement in low light conditions, however the improvements were minimal in normal lighting conditions.

Keywords: bio-inspired, thesis, object classification, object detector, computer vision, TMO, HDR, LDR, annotation, improve low light image, low light, blowfly

Subject: Engineering thesis

Thesis type: Masters
Completed: 2021
School: College of Science and Engineering
Supervisor: Dr. Russell Brinkworth