In the field of aquaculture production, fish density detection is a key link to production management. In view of the clustering of underwater fish, the technology based on Congested Scene Recognition Convolutional Neural Networks (CSRNet) is used to combine the VGG-16 without the full connection layer with the hollow convolutional neural network, which maintains the resolution and expand the perceptual domain, thus generating a high-quality map of the distribution density of fish. The results show that the detection accuracy of CSRNet in the simulated fish data set is over 90%. The predicted density map is similar to the real situation with a small distortion. It also performs well in predicting the real fish density. Compared with the traditional raster image-based method, CSRNet improves the accuracy by approximately 10%. In the meantime, CSRNet performs better than Faster R-CNN, which is also based on VGG-16. The research shows that the software of detection system constructed can detect in real time whether the fish density in the fixed point area is in the normal range. It is beneficial to prevent the phenomenon of high density hypoxia of fish, increase the output of fish, and achieve intelligent aquaculture.