Towards monitoring of a cricket production using instance segmentation

A growing world population requires sufficient food to sustain itself. Therefore, increasingly more resources are required to produce the food. Insects are a viable food and feed alternative since their production requires only a fraction of the resources that conventional livestock needs. For the efficient production of insects, automation technology is needed. An automatic monitoring of the insects’ growth ensures stable production processes and a high product quality. The use of a camera with image processing using neural networks makes it possible to detect insects, measure their features such as shape and colour and enables to derive their age, size, and health. In this paper, instance segmentation using mask scoring regional convolutional neural network (Mask Scoring R-CNN) shows good results in detecting house crickets (Acheta domesticus). A dataset is created consisting of six images, showing 1,022 insect instances, of a real-world cricket production facility to train and test the algorithm. Furthermore, image augmentation by cropping, flipping and rotating is applied to the set to solve the problem of limited data. By combining the augmentations, 288 different trainings are compared to find the best augmentation strategy. The evaluation of the algorithm uses two variations of the F1-score: one variation to estimate the capabilities of producing qualitative segmentation masks and another to estimate the detection capabilities. For the estimation of the detection capabilities, a rule termed ‘centre over ground truth’ is developed. The results show that the presented method is suitable for monitoring a cricket production facility with a recall of 76.6% and a precision of 96.2%.