การแบ่งส่วนภาพปลานิลสองขั้นตอนโดยใช้การเรียนรู้อย่างลึกด้วยเครือข่ายคอนโวลูชันพีระมิดเชิงพื้นที่

บทคัดย่อ

Tilapia is one of the most important fish in the aquaculture industry of Thailand. This research proposes a method for creating a Tilapia Image Segmentation model using deep learning with spatial pyramid convolutional network based YOLOv8 to identify the pixels of Tilapia swimming in turbid water. The process involves creating a set of images of Tilapia in water using an underwater camera. The training process of Tilapia Image Segmentation is done using Supervised Deep Transferred Learning. The structure of the YOLOv8 model is elucidated and discussed. The performance of Tilapia Image Segmentation based on YOLOv8 is evaluated and compared to a Mask R-CNN model. The results show that under a small-sized image dataset, the YOLOv8 model has better accuracy in determining Tilapia pixels from various perspective images than the model-based Mask R-CNN method, with an accuracy improvement of up to 6% on average. The YOLOv8 model has an Area Under the Curve (AUC) value of 0.988, indicating its ability to identify Tilapia pixels in images with turbid water and unclear fish shapes. This method does not require any image quality enhancement processes, which benefits the reduction of complexity in development, while still accurately identifying Tilapia pixels. Moreover, using the small-sized Tilapia-image dataset in this method reduces data preparation costs and low computational complexity.