A Novel Approach for Detection and Classification of Fish Species

Fish categorization and identification in the wild continue to be challenging because of a number of characteristics connected to manual labor, which leads to inferior or false reporting. Without direct interaction with humans, the camera is used by the proposed deep neural fish classification system to detect fish automatically. This research work has produced a You Only Look Once (YOLO) architecture-based method for classifying fish species. YOLO is a neural network-based object detection technique that will only be used once. Unlike the generic convolutional networks, which require thousands of networks and sufficiently long computing times to generate an image. In the first phase of development the proposed classification system focuses on the variation in fish scale and position and seeks out distinguishing traits to differentiate fish. The fishes are detected, positioned correctly, and fed to classifiers. Assuming that the sorts of fish captured are related to the surrounding environmental conditions, such as a boat, fishing gear, daylight, and weather, to infer the type of fish the second phase uses context information from the scenario is used. The final prediction is then estimated with the help of averaging and reweighting the two phases of the initial categorization predictions. It aids in determining the accuracy and producing suggestions tailored to particular species. Images of gilt head bream, trout, red sea bream, red mullet, sea bass, striped, red mullet, shrimp, horse mackerel, and black sea sprat from nine distinct classes are featured in the dataset utilized in this research. Underwater classification presents difficulties including background noises, image distortion, the existence of other water bodies in photos, image quality, and occlusion, the majority of algorithms now in use concentrate on classifying fishes outside of water. Compared to the previously suggested ways, the proposed strategy guarantees a significant improvement in discrimination accuracy. The proposed method outperforms the existing models by 6% and achieves an accuracy of 99.89%. This strongly implies that the approach would still work with a larger dataset.