Masking technique based attention mechanism for off-type identification in plants

Abstract

Off-type plants, or other distinct varieties of the same plant, need to be detected and removed at early stage from the fields to preserve the genetic integrity and significant traits of a given plant variety. At present, these off-type plants are manually identified in the cultivation fields through morphological analysis of leaves. This method is time-consuming and involves skilled labor to define the off-type plant. Hence, masking technique based attention mechanism is proposed for identifying the off-type from the leaf images captured in the cultivation fields. The proposed method recognizes the plant variety by locating and visualizing the prominent parts of the leaf which is similar to the human visual attention mechanism. When the prominent parts of the leaf image are masked, there is a significant drop in the class probability of the convolutional network. This reflects the importance of such prominent parts which directly account for its variety. An attention module is designed using neural networks which focuses on such significant regions and improves the performance of the network. The proposed method is tested on a field image dataset consisting of 1235 images of Sunflower leaves and achieves the highest mean accuracy of 98.25% in identifying the plant variety.