Weight information plays a pivotal role in sturgeon breeding and production management. However, manual measurement is time consuming and labor intensive due to the immense size of the sturgeon. Due to the unique body shape of the sturgeon, traditional image segmentation algorithms struggle to extract the necessary features from sturgeon images, which makes them unsuitable for this particular species. Moreover, accurately measuring weight in an occlusion environment is difficult. To address these challenges, an improved YOLOv5s model with a context augmentation module, focal-efficient intersection over union, and soft non-maximum suppression was proposed in this paper. To validate the model’s feasibility, the improved YOLOv5s model was first pre-trained using the sturgeon dataset, followed by further training on the occlusion dataset for segmentation tasks. Based on the phenotypic data obtained from the improved model, a multilayer perceptron method was used to estimate the sturgeon’s weight accurately. Experimental results demonstrated that the average precision of the improved YOLOv5s model reached 89.80% under occlusion conditions, and the correlation coefficient of noncontact weight measurement results reached 89.80%. The experimental results showed that the improved algorithm effectively performs segmentation of sturgeon in occlusion conditions and can accurately estimate the mass.