Monitoring the growth of insect larvae using a regression convolutional neural network and knowledge transfer

Abstract

Recently, there has been an increase in the popularity of breeding insect larvae (Tenebrio Molitor and Hermetia Illucens). Dimensioning larvae and observing their growth over time is a key component of monitoring insect larvae breeding. Due to the high number of larvae in the analysed images (dense scenes) and their overlap, determining the size distribution of larvae in real-time is a research challenge. In this work, we proposed an efficient method for determining the size distribution of larvae based on a regression convolutional neural network (RegCNN) and knowledge transfer. Larval width was chosen as the main measured larval parameter due to its ease of registration in dense scenes. The larval length L and its volume V were determined indirectly using determined regression models L(width) and V(width). RegCNN training was performed using knowledge transfer to omit the time-consuming labelling of multiple images containing larvae at different growth stages. Training used quartiles (lower quartile, median, upper quartile) of larval widths determined using improved multistage larvae phenotyping based on classical computer vision methods and larvae segmentation model. Finally, our approach required labelling only a few images for calibration purposes. The study evaluated different RegCNN architectures: pre-trained on ImageNet (ResNet, EfficientNet) and custom with a reduced number of model parameters. The proposed method was validated for the distribution of larvae characterised by width quartiles taking values from 1.7 mm to 3.1 mm, corresponding to an average larval length of 16 mm to 28 mm. For the best evaluated model (ResNet18) in larval width estimation, we obtained RMSE = 0.131 mm (average RMSE = 1.12 mm for larval length estimation) and R2=0.870 (coefficient of determination) with an average inference time of 0.30 s/box. The best proposed custom architecture (TenebrioRegCNN_v3) achieved slightly lower accuracy (RMSE = 0.134 mm, R2=0.864) with about five times lower inference time per image than ResNet18. The quantitative results confirmed the proposed method’s potential to be applied in real breeding conditions.