Andmebaasi logo
 

Importance of mosaic augmentation for agricultural image dataset

Laen...
Pisipilt

Kuupäev

2024

Kättesaadav alates

Autorid

Kodors, S.
Sondors, M.
Apeinans, I.
Zarembo, I.
Lacis, G.
Rubauskis, E.
Karklina, K.

Ajakirja pealkiri

Ajakirja ISSN

Köite pealkiri

Kirjastaja

Estonian University of Life Sciences

Abstrakt

The yield estimation using artificial intelligence is based on object detection algorithms. Firstly, the object detection algorithms identify the number of fruits on images, then tree fruit load is predicted using regression algorithms. YOLO is a popular convolution neural network architecture for object detection tasks. It is sufficiently well studied for fruit yield estimation. However, the experiments are traditionally restricted to only one specific fruit category and growing season. This is a big shortcoming for the smart solutions like agro-drones, which must automatically complete yield monitoring of the most popular fruit species in commercial orchards. Therefore, the modern studies related to yield estimation increasingly raise attention to multi-stage, multi-state and multi-specie detection tasks. The multi-stage datasets can be described as a collection of multiple sub-datasets, e.g. flowers, fruitlets and fruits. The multi-state dataset can contain classes like mature, immature or damaged fruits. Meanwhile, the multi-specie dataset contains images with representatives of multiple cultures. However, if classic object-detection tasks like urban or indoor object detection have multiple classes presented in one image, then yield estimation datasets usually have images with only one class presented on them. Therefore, an image shuffle or mosaic augmentation are the intuitive training strategies of YOLO for object detection working with a collection of multiple single class datasets. We applied the YOLOv5m model to test both strategies, which were verified on three datasets: apple fruits (MinneApple), pear fruits (Pear640) and pear fruitlets (PFruitlet640). Our experiment showed that mosaic augmentation improves mAP@0.5:0.95 better than simple image shuffle. The mean difference between both strategies is equal to 0.0438.

Kirjeldus

Received: August 24th, 2023 ; Accepted: February 6th, 2024 ; Published: February 23rd, 2024 ; Correspondence: sergejs.kodors@rta.lv

Märksõnad

augmentation, deep learning, yield estimation, object detection, precision horticulture, articles

Viide

Kollektsioonid