Focusing on the industrial development needs of accurate underwater target recognition in aquaculture, and aiming at the problems of low target recognition accuracy of small samples and poor adaptability of model algorithm to scenarios, a small sample aquaculture fish recognition method based on improved cycle constraint adversarial network (CycleGAN) sample amplification and attention enhancement transfer learning was proposed. Firstly, the underwater sampling equipment was used to collect the images of the actual and controllable breeding scenes of Larimichthys crocea, and the controllable scene images were used as the auxiliary sample set. CycleGAN was used as the basic framework to realize the migration of auxiliary samples to the actual breeding scene images. In particular, an optimization method of the loss function of the migration model based on the maximum mean discrepancy (MMD) was proposed. Then in the transfer learning phase, ResNet50 was used as the basic framework, and SK-Net (selective kernel network) attention mechanism optimization model was introduced to improve the perception ability of different receptive field targets, so as to improve the recognition accuracy of the model for unconstrained fish targets. The experimental results showed that the method proposed effectively improved the recognition ability of fish small sample targets, with a recall rate of 94.33% of fish recognition, and an mAP of 96.67%, providing effective technical support for the next step of fish behavior tracking and phenotype measurement.