Abstract: To estimate the quantity of Trachinotus ovatus in marine cages accurately, a method for estimating the quantity of fish is proposed by using high-frequency horizontal mechanical scanning sonar and deep learning technology. Differentiating water layers and clustering layer by layer to realize counting is the main way of this method, which mainly involves three parts: fish identification, fish cluster and fish quantity fitting. Firstly, high-frequency horizontal mechanical scanning sonar is used to conduct spiral detection on marine cages to obtain fish image information, which is labeled to build training data set of improved CS-YOLOv8s. After training, the CS-YOLOv8s model is used to recognize fish location information in the images. Secondly, the cages are divided into water layers with a water depth spacing of 40 cm, and the identification coordinate data of each water layer are clustered through DBSCAN method to generate fish quantity data of each water layer. Finally, the quantity data of each water layer is fitted with the given quantity of fish in the cage, and the fitting model of fish quantity is established. The results show that in the quantitative experiment of marine cages, the accuracy of this method is 87.14%, and it can achieve a good estimation of the quantity of T. ovatus.