TRL-Net: a tiny object recognition and localization method in the artemisinin manufacturing process

Abstract

Abstract Artemisinin is a key pharmaceutical ingredient for dysentery treatment. In the manufacturing process of artemisinin, artemisinin recognition and localization (ARL) is an important step to guarantee high drug purity. As an image processing problem, real-time solvent volatilization images of artemisinin production are used to determine the recognition of artemisinin materials. Images with artemisinin have small and intensive properties, which increases the difficulty of identification and location. Therefore, this paper proposes a tiny recognition and localization network (TRL-Net) based on a region-based convolutional neural network (R-CNN) to improve the performance of ARL. In TRL-Net, we establish a deep extraction backbone network with specially designed tiny necks to catch detailed features. Furthermore, tiny cross-entropy and Smooth-L1 loss functions are discovered to reduce the severe influence of negative samples on locating actions. Finally, experimental results on the real-world artemisinin dataset indicate that our proposed approach outperforms other compared methods.