Abstract
Detecting defects such as holes, biofouling, and vegetation in fishnets is crucial for efficient and sustainable fish farming (Aquaculture, Source) [1]. The LIACi (Lifecycle Inspection, Analysis, and Condition information system) project focuses on enhancing underwater ship and fish farm monitoring by integrating AI/ML-based automation features [2]. In this poster, we illustrate our approach to detecting defects in fishnets by employing computer vision and deep learning techniques.
We assembled a comprehensive annotated image data set capturing diverse fish- net defects, including holes, marine growth (biofouling), and uncertain defects. The data set served as the foundation for training our deep learning model. However, un- derwater conditions present challenges such as variable lighting, water turbidity, and image distortions, impacting image quality. Additionally, we encountered challenges related to the availability of images with improper resolution, camera angles, or light- ing conditions during capture. Addressing these challenges was essential to ensuring the model’s reliability.
To evaluate our model’s effectiveness, we extensively tested it on a diverse fish farm video data set. The results were promising, with our deep learning model achieving an F1 score of approximately 70%. A few examples of precisely identified defects are given in Figure 1. Performance can be further improved by expanding the training data set with proper image sets. This signifies significant progress in automating defect detection in fishnets. A demonstration will be available at https://liaci.sintef.cloud.
We assembled a comprehensive annotated image data set capturing diverse fish- net defects, including holes, marine growth (biofouling), and uncertain defects. The data set served as the foundation for training our deep learning model. However, un- derwater conditions present challenges such as variable lighting, water turbidity, and image distortions, impacting image quality. Additionally, we encountered challenges related to the availability of images with improper resolution, camera angles, or light- ing conditions during capture. Addressing these challenges was essential to ensuring the model’s reliability.
To evaluate our model’s effectiveness, we extensively tested it on a diverse fish farm video data set. The results were promising, with our deep learning model achieving an F1 score of approximately 70%. A few examples of precisely identified defects are given in Figure 1. Performance can be further improved by expanding the training data set with proper image sets. This signifies significant progress in automating defect detection in fishnets. A demonstration will be available at https://liaci.sintef.cloud.