Toward in situ zooplankton detection with a densely connected YOLOV3 model

Zooplankton play an important role in the global marine carbon cycle, and as a useful indicator of aquatic health, the distribution and abundance of zooplankton organisms could provide early warning for natural disasters. With the rapid development of the observation sensors and platforms, many advanced detection methods such as deep neural networks are pursued to realize the in situ and autonomous zooplankton observation. However, the features of zooplankton might be lost in the deep neural network transmission due to both convolution and downsampling operations, especially for the subtle features which are critical in the identification of the zooplankton taxonomic group. Therefore, this paper proposed an improved YOLOV3 model with densely connected structures to improve the reusability of the features during transmission in the model. The experiment results demonstrate the performance of the proposed method is more suitable for the in situ zooplankton detection by comparing it with other state-of-the-art models.