Automatic shrimp counting method using local images and lightweight YOLOv4

Shrimp counting is a fundamental operation for biomass estimation in shrimp culture. It is also vital for achieving reasonable feeding and improving breeding efficiency. The application of computer vision technology in the counting of aquatic products is nondestructive and highly efficient. However, owing to the small size, transparent body, and complex background of shrimp in the actual culture environment, existing methods cannot ensure the lightweight of the model applied while satisfying the counting accuracy, and it is difficult to achieve accurate real-time shrimp counting. Therefore, an automatic shrimp counting method using local images and lightweight YOLOv4 (Light-YOLOv4) was proposed in this study. Multiple local shrimp images were randomly cropped from the original top view images previously collected using image processing technologies to construct a counting dataset. Subsequently, a local shrimp counting model based on Light-YOLOv4 is constructed and trained using transfer learning. Based on the trained model, the number of shrimp in each local shrimp image was predicted. The number of shrimp in the original shrimp image was obtained through a merging process, and the number of shrimp in the culture area was determined using the frame average method. The method was tested on a real shrimp dataset, and the Light-YOLOv4 local shrimp counting model achieved a counting precision of 92.12%, recall of 94.21%, F1 value of 93.15%, and mean average precision of 93.16%. Compared with other counting models, the proposed method exhibits a better comprehensive performance in terms of the counting accuracy, model size, and detection speed. Furthermore, when shrimp were counted within the entire culture area, the results were consistent with the true values. (C) 2022 IAgrE. Published by Elsevier Ltd. All rights reserved.

Automated summary

This study proposes an automatic shrimp counting method using local images and lightweight YOLOv4. The method utilizes a counting dataset constructed from cropped local shrimp images and trains a model based on transfer learning. The results show that the method achieves accurate shrimp quantity prediction and outperforms other models in terms of counting accuracy, model size, and detection speed.