SLCOBNet: Shrimp larvae counting network with overlapping splitting and Bayesian-DM-count loss

Abstract

Estimating the number of shrimp larvae plays a critical role for achieving reasonable feeding in aquaculture. However, previous shrimp larvae counting models failed to accurately distinguish between shrimp larvae and other objects due to the loss of shrimp larvae continuity information. Also, shrimp larvae counting has the challenges of multi-scale changes, transparent body overlap, and background noise. To solve the above problems, a novel shrimp larvae counting network called SLCOBNet is proposed. First, overlapping splitting image is devised, with each patch sharing half of its area, ensuring the continuity of information regarding shrimp larvae between patches. Then, shrimp larvae feature of different scales and multi-scale density maps are obtained through the feature pyramid aggregation and the regression head with multi-scale receptive fields, respectively. Moreover, a novel loss function called Bayesian-DM-Count loss, composed of DM-Count loss and Bayesian loss, was designed to address the existing transparent body overlap and background noise problems. Experiments were performed by collecting shrimp larvae data from the breeding trays of a real aquaculture farm. The extensive experimental results on three shrimp larvae datasets have shown that SLCOBNet achieves 3.27, 3.61 and 1.28 in Mean Absolute Error. Hence, the proposed method exhibits a better counting accuracy compared to other counting methods. Furthermore, the predicted results were consistent with the true values.