Counting shrimp larvae is a tedious task for both hatcheries and shrimp farmers as it is time-consuming and requires intensive manual labor, taking about 15 to 20 minutes to count 500 to 700 larvae.
A team of researchers from the South China Agricultural University and the South China Sea Fisheries Research Institute (CAFS) has developed an automatic method for counting Penaeus monodon larvae using keypoint regression to obtain the number of larvae in images captured solely by smartphones.
The main findings of the study are summarized as follows:
- They collected a large number of Penaeus monodon larvae images without any additional assisting instruments and built a dataset called Penaeus_1k.
- A counting method for P. monodon larvae that can accurately predict the number of larvae in the image and precisely locate the keypoints of the larvae in the image.
- Extensive experiments were conducted on the Penaeus_1k dataset to demonstrate the effectiveness of the proposed method.
Methods for Counting Shrimp Larvae
Methods for counting shrimp larvae can be classified into traditional estimation methods and image processing methods.
Traditional estimation methods include manual methods and fish methods. In the manual method, thousands of larvae are separated into numerous containers and counted in batches by different individuals. This process is time-consuming, labor-intensive, and stressful for the larvae due to prolonged exposure.
In the fish method, a group of larvae is weighed and divided by the weight of a single larva. This method requires the larvae to be taken out of the water, which is dangerous for their health.
Image Processing Methods
Image-based methods for counting shrimp larvae can be classified into traditional and deep learning-based approaches.
The traditional method relies on image processing techniques such as thresholding, erosion, dilation, and connected component analysis to segment and locate shrimp larvae in images by exploiting the salient features of larval hepatopancreas.
Deep learning-based approaches utilize target detection and instance segmentation techniques to automatically determine the count of shrimp larvae from static images. However, accurately locating the larvae is challenging due to their overlap and adherence.
Collection of Shrimp Larvae Data
To standardize the process of collecting shrimp larvae data, the researchers used a container with an appropriate volume of water. They added a small number of larvae to the container as the first group, followed by seven additional groups, each containing successively more larvae.
To ensure consistency, five individuals took turns to capture 50 images after each larval addition. The image capturer held the mobile phone at a distance of 300mm to 500mm above the container and captured images that included all the larvae in the container.
In order to obtain a comprehensive database, the images were collected from different perspectives, resulting in variations in background, perspective, and light intensity.
Smartphone Application for Shrimp Larvae Counting
The algorithm proposed by the researchers for automatic shrimp larvae counting was implemented through a WeChat applet, which provides a concise and convenient user interface.
On the applet’s home page, users can click on the green counting button located in the middle. They have the option to either capture a photo or select a shrimp larvae image from their smartphone’s album.
Subsequently, the user can upload the image to the cloud server for processing. On the cloud server, the mask R-CNN algorithm is initially employed to extract the container from the image, thus removing background interference.
Finally, the results of the shrimp larvae counting, along with the processing time, the original image, and the labeled image, are sent to the results page for visualization. Each counting result is saved in the counting history for future review.
Features of the Application
“In this study, we propose an efficient shrimp counting method based on PLCS and HRNet-w48. The proposed method roughly locates the keypoints of shrimp larvae while simultaneously obtaining the number of shrimp larvae in the images,” report the researchers.
They highlight that, unlike traditional image processing methods, it does not require the use of additional equipment to ensure consistency in the image acquisition environment.
However, they emphasize that the counting accuracy by the proposed method can be significantly affected by the image quality.
“To facilitate automatic shrimp counting, we developed a system based on deep learning methods and smartphones. The system uses smartphones to capture images of shrimp larvae, which are then uploaded to a server for processing to count the number of shrimp larvae in the images,” they noted.
“This study proposes an easy-to-use, efficient, and accurate counting method for black tiger shrimp larvae. This method can accurately estimate the number of larvae in the container, improve the efficiency of transactions of black tiger shrimp larvae between hatcheries and shrimp producers, and help breeders have a more accurate understanding of the number of shrimp larvae purchased,” conclude the researchers.
The proposed method was evaluated on the Penaeus_1k test dataset, achieving favorable counting results.
“The average accuracy rate reached 93.79%, with an MAE of 33.69 and an MSE of 34.74. Compared to density map methods, our proposed approach demonstrates overall superior performance. It not only accurately determines the key positions of the larvae but also exhibits high precision and stability,” they concluded.
References (open access)
Li, Ximing, Ruixiang Liu, Zhe Wang, Guotai Zheng, Junlin Lv, Lanfen Fan, Yubin Guo, and Yuefang Gao. 2023. “Automatic Penaeus Monodon Larvae Counting via Equal Keypoint Regression with Smartphones” Animals 13, no. 12: 2036. https://doi.org/10.3390/ani13122036