LAUSR

Abstract Manual labor involved in shrimp extraction selection accounts for an extremely high proportion of processing time and also entails reduced accuracy and efficiency moreover even it could induce potential safety hazards. The key to substitute the manual process with automation lies in the identification and pinpointing of the penultimate joint in shrimps. Therefore, a cascaded neural network is proposed in this study to implement the detection of key points in a multi-shrimp scenario processing. More specifically, our model includes two stages: a shrimp detector based on YOLOv3 and followed by a pose estimator based on Convolutional Pose Machine (CPM). With the combination of attention mechanism and improved NMS strategy, our detector is equipped to resist noise interference in dense case, ubiquitous on the production line. Experimental results indicate that both the detection rate and the speed information extraction have achieved the standard of industry applications.