Robust Leaf Detection using Shape Priors within Smaller Datasets

Debojyoti Misra, Tushar Sandhan

Abstract: Our brain can process visual information, which helps understanding the shape of an object. If never seen before, an accurate description of the shape can help ease the task for a human who is looking for an object. Data scarcity in agriculture is primarily due to the labor-intensive and cost-intensive nature of collecting, as well as the requirement for expertise to label them. Our task of leaf detection has only one kind of object, which has some general shape features. To make the task of learning easier from a comparatively smaller dataset, we automatically learn shape prototypes from leaves and use them as templates to generate shape-specific features to incorporate prior knowledge into the neural network. We use this method to generate prototypes from the Plant Village dataset and use them for detection in the Plant-Doc dataset to improve the mean average precision (mAP) by 3% over the state-of-the-art Faster-RCNN model. These kinds of experiments show the cross-dataset generalizability of the proposed method.

External IDs:dblp:conf/icpr/MisraS24