Abstract: The accurate statistics of the number of turbot (Scophthalmus maximus) eggs is a crucial factor affecting the seedling selection process. Due to the small size, high transparency, and easy adhesion of turbot eggs, the manual counting methods are inefficient and prone to significant errors. To achieve automated, rapid and accurate counting of turbot eggs, a turbot fish egg counting method based on convolutional block attention mechanism and U-shaped convolutional neural network (CBAM-UNet) is proposed. According to the imaging characteristics of turbot eggs, a standardized fish egg sampling structure composed of an industrial camera, a diffuse reflection light source, and an image acquisition box was first designed and developed to obtain shadow-free high-definition egg images and construct a fish egg image sample set. Then, with UNet network as the basic semantic segmentation model, a dual attention mechanism was further introduced to enhance the segmentation of the boundaries and details of fish eggs and the background, so as to improve the expression ability of the model to the egg features and the segmentation accuracy. Finally, a multiple linear regression model was constructed based on the segmented fish egg area, shooting height, and the number of fish eggs to achieve accurate counting of the fish eggs. Experimental results show that the proposed recognition and counting method based on CBAM-UNet can effectively improve the accuracy of fish egg counting, with an average counting error of 6.32%, lower than the average counting error of other models and manual statistical methods (Quality comparison method).