Abstract: Shrimp larval counting is a fundamental operation for biomass estimation in shrimp farming and sales, and it is crucial for aquaculture management and yield enhancement. Due to the factors such as small size of shrimp larvae, significant occlusion, and varying density, the current automated counting methods are difficult to achieve accurate larval counting at certain densities. To address this issue, we propose a shrimp counting (SC) model for shrimp larval counting which combined frontend and backend counting network based on Bayesian probability distribution. This model primarily consists of a frontend module, an attention module and a backend module. The frontend module first extracts discriminative phenotypic features, while the attention module reorganizes these features to enhance local attention to the images. Secondly, the backend module generates a predicted probability density map for shrimp larval distribution. Thirdly, the Bayesian loss function is utilized to adjust the model parameters and improve the accuracy of shrimp larval counting. To validate the effectiveness of the proposed method, we constructed a shrimp larval counting dataset with two different density conditions and conducted multiple experimental comparisons on the dataset. The overall accuracy reached 93.325%, with mean absolute error and mean squared error being 96.304 and 154.567, respectively. Compared with the current mainstream counting methods Black-Litterman (BL), Contextual Scene Recognition Network CSR-Net), Boosting Crowd Counting via Multifaceted Attention BCCMA, the proposed model exhibits the highest accuracy and the lowest loss. It applies to automated shrimp larval counting in hatcheries, sales and stocking scenarios.