Using Agents and Unsupervised Learning for Counting Objects in Images with Spatial Organization

Jacopin, Eliott; Berda, Naomie; Courteille, Léa; Grison, William; Mathieu, Lucas; Cornuéjols, Antoine; Martin, Christine

doi:10.5220/0010228706880697

Cited by 3 publications

(1 citation statement)

References 9 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Most of them belong to the Object Based Image Analysis (Josue Nahun Leiva et al, 2017; Koh et al, 2019; Torres-Sánchez et al, 2015; Varela et al, 2018; Zhao et al, 2018). The identification process can be done based also on the expert knowledge (Gnädinger and Schmidhalter, 2017; Jacopin et al, 2021; T. Liu et al, 2016) or by calibrating a statistical model over a training dataset (Calvario et al, 2020). More recently, approaches based on deep-learning (DL) have been proposed.…”

Section: Introductionmentioning

confidence: 99%

Plant detection and counting from high-resolution RGB images acquired from UAVs: comparison between deep-learning and handcrafted methods with application to maize, sugar beet, and sunflower

David

Daubige

Joudelat

et al. 2021

Preprint

View full text Add to dashboard Cite

Plants density is a key information on crop growth. Usually done manually, this task can beneficiate from advances in image analysis technics. Automated detection of individual plants in images is a key step to estimate this density. To develop and evaluate dedicated processing technics, high resolution RGB images were acquired from UAVs during several years and experiments over maize, sugar beet and sunflower crops at early stages. A total of 16247 plants have been labelled interactively. We compared the performances of handcrafted method (HC) to those of deep-learning (DL). HC method consists in segmenting the image into green and background pixels, identifying rows, then objects corresponding to plants thanks to knowledge of the sowing pattern as prior information. DL method is based on the Faster RCNN model trained over 2/3 of the images selected to represent a good balance between plant development stage and sessions. One model is trained for each crop. Results show that DL generally outperforms HC, particularly for maize and sunflower crops. The quality of images appears mandatory for HC methods where image blur and complex background induce difficulties for the segmentation step. Performances of DL methods are also limited by image quality as well as the presence of weeds. An hybrid method (HY) was proposed to eliminate weeds between the rows using the rules used for the HC method. HY improves slightly DL performances in the case of high weed infestation. A significant level of variability of plant detection performances is observed between the several experiments. This was explained by the variability of image acquisition conditions including illumination, plant development stage, background complexity and weed infestation. We tested an active learning approach where few images corresponding to the conditions of the testing dataset were complementing the training dataset for DL. Results show a drastic increase of performances for all crops, with relative RMSE below 5% for the estimation of the plant density.

show abstract