Estimating fish populations is vital for fisheries management and ecosystem conservation. Sonar imaging-based estimation is a promising method because it is unaffected by murky waters and low costs. This study proposes a double transfer learning approach for training deep learning models to improve the accuracy of fish counting in sonar images. In the proposed method, deep learning models are sequentially trained on ImageNet and DeepFish datasets before being trained on Sonar dataset to leverage domain similarities. Experiments are carried out with VGG16, ResNet50, RegNetX, and RegNetY models to verify the effectiveness of the proposed method. By applying the proposed method, RegNetY reduces mean absolute error (MAE) by 20% and VGG16 reduces root mean square error (RMSE) by 22% for fish counting in sonar images compared to these models without DeepFish knowledge.