$Data centric approach for oil palm trees detection$
Keywords: data-centric AI, model-centric AI, objects detection, deep learning, Artificial Intelligence
TL;DR: Implementation of oil palm trees detection model in UAV images using the data centric AI approach.
Abstract: To enable African agritech startups to keep up with the progress made in recent years in the field of artificial intelligence, we decided to set up a tool for automatically counting oil palm trees in drone images. This tool is a continuation of digital africa’s commitment through its data for digital africa (D4DA) program, which aims to promote the use of data by African agritech startups, and is also an improvement on a first solution obtained following a challenge organized on the zindi platform. This first solution was based on a regression approach to predict the number of palm trees in images, without necessarily saying where these palm trees are actually located in the image a sort of black box, as we like to say. The new solution proposed consists in detecting the palm trees in the images for counting purposes, with a view to making the model more explicable. This solution was implemented using a data-centric AI approach with the faster r-cnn pre-trained model. Indeed, the faster r-cnn model was fine-tuned, adapting the output of the last layer to the number of objects we wished to detect in the images, in this case a single object (the oil palm). It should also be noted that with faster r-cnn, we add +1 to the number of objects to define the total number of classes, this +1 corresponding to the background class. The results obtained in this study are quite encouraging:
with 312 images annotated on the principle of data centric AI, which we divided into train-test (90%,10%) for training, we obtained an average precision (AP) of 77% at the threshold of 0.75 IoU on the test data. These results can be considerably improved on the basis of data-centric AI principles. Our main aim in this study was to demonstrate the interest of this new approach in the field of data science in
general and computer vision in particular.
Submission Category: Machine learning algorithms
Submission Number: 15
Loading