Accurately obtaining the number of fish is a fundamental process for biomass estimation in fish culture. It not only helps farmers calculate the reproduction rate and estimate the production potential accurately but also serves as a guide for survival rate assessment, breeding density control, and transportation sales management. It can be said that fish counting runs through multiple links such as breeding, transportation, and sales. Among these links, fish live in different environments and their body size is also various, bringing certain difficulties to fish counting. Aiming at the above problems, a fish counting method based on multi-scale fusion and no anchor YOLO v3 (MSF-NA-YOLO v3) was proposed. Firstly, multi-source fish images were collected to construct a fish counting dataset with a total of 1858 images. Secondly, the feature extraction network of YOLO v3 was improved, and a feature extraction method based on multi-scale fusion was proposed to enhance the feature expression of fish images. Finally, the CenterNet was used as the detection network of YOLO v3, and then a fish target detection network based on no anchor was proposed to identify fish targets in images and realize fish counting. The collected fish counting dataset was randomly divided into a training set, validation set and test set. The training set and validation set accounted for 90% of the dataset, with a total of 1672 images, and the test set accounted for 10% of the dataset, with a total of 186 images. The ratio of the training set to the validation set was 9∶1, containing 1505 and 167 images, respectively. The MSF-NA-YOLO v3 fish counting model was trained and validated by using the transfer learning method. When the training loss and validation loss became stable, the training stopped and the best fish counting model was obtained. Based on this model, the fish images of the test set were counted and a precision of 96.26%, recall of 90.65%, F1 value of 93.37%, and average precision of 90.20% were achieved. Compared with the fish counting model based on the original YOLO v3 feature extraction method and the single scale fusion feature extraction method, the precision of the fish counting model based on the feature extraction method proposed was increased by 0.51% and 0.72%, respectively, recall was increased by 0.44% and 1.72%, respectively, F1 value was increased by 0.47% and 1.24%, respectively, and mean average precision was increased by 0.45% and 1.87%, respectively, indicating that the proposed feature extraction method had better performance. Compared with the fish counting method based on YOLO v3, YOLO v4, and ResNet+CenterNet, the recall was increased by 5.80%, 1.84%, and 3.48%, respectively, F1 value was increased by 2.26%, 0.33%, and 1.68%, respectively, and mean average precision was increased by 5.96%, 1.97%, and 3.67%, respectively. Thus, the proposed method had a good overall performance and can provide support for the realization of fishery automation and intelligence.