Fall armyworm is a pest that eats almost everything grown by humans (rice, corn, wheat, etc.) slightly rampaged in India in 2018, which led to significant damage to the crop sorghum and bajra.

For such feats, our hero was awarded special attention by researchers. This led to the fact that its presence on plants was decided to be detected automatically.

The dataset was collected from photos of various pests from the Internet. Moreover, 25 real photos were taken. After augmentation (counter-clockwise rotation, horizontal and vertical flipping, and zooming), the researchers had 798 examples.

Three tasks were solved: image classification (whether or not the image has the “wanted” worm), prediction of the bounding box (namely the rectangles surrounding the objects), and object masks (pixels corresponding to the object position). The Mask R-CNN architecture worked best. The architecture of Mask R-CNN (Region-Based Convolutional Neural Network) consists of two main parts.

The first part named Region Proposal Network is simply a Neural Network that proposes the bounding boxes for multiple objects that are available within a particular image. In the second part, in parallel to predicting the class and bounding box offset, Mask R-CNN also outputs a binary mask for each object. Thus, the Mask R-CNN architecture makes it possible to solve all three tasks simultaneously.