Active Learning for Crop-Weed Discrimination by Image Classification from Convolutional Neural Network’s Feature Pyramid Levels

Zahidi, Usman A. and Cielniak, Grzegorz (2021) Active Learning for Crop-Weed Discrimination by Image Classification from Convolutional Neural Network’s Feature Pyramid Levels. In: 13th International Conference, ICVS 2021, September 22-24, 2021.

Full content URL: https://doi.org/10.1007/978-3-030-87156-7_20

Full text not available from this repository.

Item Type:Conference or Workshop contribution (Paper)
Item Status:Live Archive

Abstract

The amount of effort required for high-quality data acquisition and labelling for adequate supervised learning drives the need for building an efficient and effective image sampling strategy. We propose a novel Batch Mode Active Learning that blends Region Convolutional Neural Network’s (RCNN) Feature Pyramid Network (FPN) levels together and employs t-distributed Stochastic Neighbour Embedding (t-SNE) classification for selecting incremental batch based on feature similarity. Later, K-means clustering is performed on t-SNE instances for the selected sample size of images. Results show that t-SNE classification on merged FPN feature maps outperforms the approach based on RGB images directly, random sampling and maximum entropy-based image sampling schemes. For comparison, we employ a publicly available data set of images of Sugar beet for a crop-weed discrimination task together with our newly acquired annotated images of Romaine and Apollo lettuce crops at different growth stages. Batch sampling on all datasets by the proposed method shows that only 60% of images are required to produce precision/recall statistics similar to the complete dataset. Two lettuce datasets used in our experiments are publicly available (Lettuce datasets: https://bit.ly/3g7Owc5) to facilitate further research opportunities.

Keywords:Active Learning, Image sampling, Robotic weeding, Convolutional Neural Network, Deep learning
Subjects:G Mathematical and Computer Sciences > G400 Computer Science
Divisions:College of Science > Lincoln Institute for Agri-Food Technology
ID Code:46648
Deposited On:16 Dec 2021 10:07

Repository Staff Only: item control page