Shrimp fry counting is an important part of the shrimp fry breeding and transaction process. Shrimp fries are small, dense and easily inactivated, which makes it difficult to count shrimp fry upon emergence. An automatic counting technology for shrimp fry with a convolutional neural network model is proposed. First, each target object in the data sample is labeled with a pixel point to obtain the true density map, then the training samples are input to the improved convolutional neural network learning to generate an estimated density map from the image features, and finally the total number of shrimp fry in the entire field of view is obtained from the density map. In order to verify the effectiveness of the method, the shrimp fry images taken in the fry farm are used as data sets and comparative experiments are conducted on different models. The results show that: compared with the classical networks such as MCNN, CSRnet and CAN, the average absolute error can be reduced by 7.6, 4.8and 3.2 respectively, and the root mean square error can be reduced by 3.35, 5.19 and 6.39 respectively. Researches have shown that this method can accurately estimate the number of shrimp fry of a certain density under a uniform backlight environment, which meets the counting requirements of shrimp aquaculture.