A Machine-Learning-Based Ultrasonic System for Monitoring White Shrimps

Computer vision has been applied in many fields recently to extract useful information from images. One such application is the monitoring of various aquaculture species to increase production. By monitoring the locations and quantities of the white shrimp, the aquaculture farmers could deliver the feed more precisely. Besides, the information about the length of the white shrimp could make the farmers estimate the delivery time accurately. Usually, to ensure the accuracy of computer vision, a relatively high-quality picture is required. However, the quality of underwater photography is easily influenced by the light, color, and turbidity of the medium. Compared to optical photographs, the imaging mechanism of ultrasound is less affected by these effects. This study proposes an innovative white shrimp monitoring system that combines ultrasound imaging technology with YOLOv4. The database of the shrimp’s ultrasound images was built by imaging the different parts of the shrimp from various viewing angles with a linear ultrasonic probe. Four models using training data from different scenarios were successfully trained to recognize the locations of the shrimps from the ultrasound images. Among these models, the best was the one trained with 650 one-shrimp, 1080 two-shrimp practical transverse ultrasound images, and 6500 images from data augmentation. Unlike optical pictures, ultrasound images are cross-sectional images that slice across certain parts of the shrimps. Therefore, a bi-directional scanning approach was adopted in this study. The ultrasound probe was carried by a two-axis motor to capture the images from two orthogonal viewing angles within a
$12.69\times12.69$
cm region of interest (ROI). After 260 images were captured, the locations of the shrimps in each image were recognized, and the centers of the bounding boxes (bandboxes) were marked. To determine the number of shrimps, a voting method was developed based on the spectral clustering algorithm to group the marked coordinates. Finally, the number, lengths, and positions of the shrimps were reported. To evaluate the accuracy of the proposed monitoring system, 100 shrimps with 40 different layouts containing one to four shrimps were tested. The averaged quantity error, position error, and length error were 97.3%, 0.97 cm, and 8.99%, respectively. The standard deviation of the position error and the length error were 0.55 cm and 5.42%, respectively. The results demonstrated that the proposed system can provide useful information on shrimps and is feasible for shrimp farming applications.