Segmentation of Fish in the Natural Habitat Using Deep Learning for Fish Size Measurement Application

Assessing the size of fish in their natural habitat could provide crucial information for managing and monitoring these resources, thereby enabling the development of sustainable fishing and resource conservation strategies. However, manual methods are time-consuming, labor-intensive, and can have adverse effects on the health of the fish. Therefore, this study proposes a novel automated monitoring method by utilizing fish image segmentation, which allows for easier estimation of their size. Our proposed model namely AquaAttSeg has Ushape architecture with an encoder and a decoder. AquaAttSeg utilizes the PVTv2-B3 model as the backbone of encoder to enhance the ability to extract semantic information from input. In contrast to conventional U-shape models that directly connect low-level features from the encoder to high-level features in the decoder, leading to information imbalance, we proposed the HFIM module, which includes two our proposed modules SAM and CAM to replace conventional skip connections in order to enhance important information from the encoder to the decoder and enhance the processing of low-level features to extract more semantic information before combine with high-level features to reduce the information imbalance. In the decoder stage, we also proposed the Convmix Transpose Custom block, which can effectively synthesize rich and multidimentional information between the encoder and decoder to efficiently recover lost information and enhance the performance of the model. The performance of our proposed model has been tested on two datasets including Deepfish dataset and SUIM dataset. Additionally, we have compared the results of our model with other state-of-the-art models, demonstrating that our model produces highly competitive and promising outcomes.