Click on “Download PDF” for the PDF version or on the title for the HTML version.
If you are not an ASABE member or if your employer has not arranged for access to the full-text, Click here for options.
Applying Generative Adversarial Networks for Sticky Paper Trap Image Generation and Object Detector Performance Enhancement
Published by the American Society of Agricultural and Biological Engineers, St. Joseph, Michigan www.asabe.org
Citation: 2022 ASABE Annual International Meeting 2200266.(doi:10.13031/aim.202200266)
Authors: Dan Jeric Arcega Rustia, Eros Allan Hacinas, Liane Angelo Acero, Lorenzo Sangco Querol, Mercedes Arcelo, Divina M Amalin
Keywords: Data augmentation, deep learning, insect counting, generative adversarial networks, sticky paper trap.
Abstract. One of the issues in training a deep learning-based object detector is the imbalance in the number of objects found in an image. This makes an object detector inadequately trained on different backgrounds and varying numbers of objects. To propose a solution to this problem, this work presents a method for generating synthetic cocoa pod borer (CPB) images and building a sticky paper trap image dataset in order to enhance the performance of a deep learning-based object detector. The proposed method was tested on sticky paper traps collected from cacao plantations with pest pressure from CPB. The sticky paper traps were scanned using a flatbed scanner while CPB images were extracted to train an auxiliary classifier generative adversarial network (AC-GAN) model and generate synthetic CPB images. It was found that the synthetic CPB images had a Frechet inception distance score of 7.14 relative to the real CPB images. The synthetic CPB images were overlaid onto the sticky paper trap training images of an object detector using various image processing methods in order to augment the number of CPBs found on each sticky paper trap image. Testing results show that the object detector trained with overlaid synthetic CPB images achieved an average precision of 0.94, outperforming the object detector trained with real overlaid CPB images which had an average precision of 0.88. This research proves the advantage of generating synthetic insect images for building and reducing the time and effort in collecting insect pest images for training deep learning models.
(Download PDF) (Export to EndNotes)