SegFishHead: A Semantic Segmentation Approach for the identification of fish species in a Cluttered Environment

Identifying fish species, in general and in a cluttered environment in particular, is not an easy task for common people without the proper knowledge of fish taxonomy. This study proposes a challenging segmentation dataset consisting of fish images collected from different live fish markets situated in West Bengal, India. A total of five freshwater fish species named, Labeo catla, Labeo rohita, Cirrhinus mrigala, Labeo bata, and Hypophthalmichthys molitrix are considered in this study. A semantic segmentation-based fish head segmentation and identification in a cluttered environment is proposed in this study. Two popular deep learning-based segmentation networks, U-Net and PSPNet, are applied with two different pre-trained backbone networks, ResNet34 and InceptionV3. Using PSPNet with a ResNet34 pre-trained backbone, a best mean IoU of 0.76 is achieved by taking the background of the image as a class label. The fishing industries as well as their stakeholders will benefit from this proposed approach in a variety of contexts.