FUnctionality Sharing In Open eNvironments
Heinz Nixdorf Chair for Distributed Information Systems

Refined Methodology for Accurately Detecting Objects from Digitized Herbarium Specimens

Title: Refined Methodology for Accurately Detecting Objects from Digitized Herbarium Specimens
Authors: Triki, Abdelaziz; Bouaziz, Bassem; Gaikwad, Jitendra
Source: https://icei2018.uni-jena.de/
Place: Jena, Germany
Date: 2018-12-20
Type: Poster

There are global initiatives such as TRY trait database that are making efforts to minimize the paucity of functional trait data. The herbarium specimens provide valuable information on functional traits including length, width, and size of leaves and petiole length. Traditionally, scientists extract such information manually, which is time consuming and prone to errors.

To overcome these limitations, worldwide scientists are applying computer vision techniques to automatically extract trait data from digitized specimen images. However, to extract the trait values pre-requisite is to accurately detect objects present on the specimen images. In the Managing Multimedia Data for Science (MAMUDS) project, we have refined a deep learning technique, which is efficiently able to detect objects from the specimen images provided by the herbarium Haussknecht, as compared to other state-of-the-art techniques.

The herbarium Haussknecht in Germany provide access to more than 30000 scanned type specimen images to researchers and public. However, the varied placement and diversity of objects such as plant specimen, scale bar, color pallet, specimen label, envelope, barcode and stamp on the specimen sheet make the task of automatic detection challenging. To address this challenge, we have developed a refined methodology RefYOLO, which is based on the pioneer object detection system called You Only Look Once (YOLO). The refinement was done by editing the activation function of VGG16 model using the Parametric ReLu (PReLU). In our study, we observed that YOLO and other state-of-the-art techniques such as Region-based Convolutional Network (RCNN) and its variants are unable to robustly detect objects from digitized specimen images. RefYOLO efficiently localizes object classes located in specimen images and scales by encoding contextual feature information about classes such as shape, contours and their appearance. After training, the performance of RefYOLO was enhanced using Average-Max pooling method (AM-PM). Overall, RefYOLO provides higher detection accuracy and need less processing time.

URL: https://doi.org/10.22032/dbt.37908