Underwater Image Based Fish Recognition

The study of fish plays an important role in a wide range of natural science researches such as fishery planning,marine ecosystem study and environmental research.The first and most critical step in fish study is fish species recognition.Traditional approaches usually rely on manual image capturing and diving-based observation to estimate the number of fish and fish species in certain water areas,which is then used to estimate key metrics such as biodiversity and distribution.In recent years,with the development of underwater sensing and imaging technologies,marine organizations from multiple countries have collected and released a large number of underwater fish images.Based on these data,researchers have proposed many machine learning models and methods to train fish recognition models.However,most of the existing works suffer from obvious limitations,as these methods are mostly based on traditional machine learning models and small-scale datasets(or species),thus can hardly be applied to train large-scale,highly accurate model.This is largely because underwater images are extremely blurry,dark and often contain a large amount of background noise and blocking objects.These extremely noisy underwater images pose a huge challenge for underwater fish recognition.This thesis focuses on the task of underwater fish recognition and conducts the following four key studies toward the goal of developing large-scale and high performance fish recognition models:(1)Deep learning based underwater fish recognition model and loss function study.This part of work systematically studies the performance of a set of deep learning models and robust loss functions for underwater fish recognition,and determines the optimal deep learning model structure and loss function for large-scale fish recognition.Based on this,this thesis also provides a series of key findings that can serve as an experimental foundation for future research.(2)Mixture contrastive learning based underwater fish recognition.To solve the problem that the noisy background in underwater images could greatly hinder the learning of the fish object,this part of our work proposes a novel Mix Contrast learning method to help the model differentiate fish object from the background by generating a purebackground image then contracting it against the original images(via contrastive leaning).The results on real-world underwater datasets show that this method can significantly improve the recognition accuracy of existing methods.(3)Discriminative feature learning based underwater fish recognition.Based on findings obtained in(2),this part of work proposes a more advanced robust training method based on discrimination feature learning and attention suppressing techniques to encourage the model to pay more attention to the fish object,and at the same time,suppress its attention to the background.This approach can greatly improve the model's robustness to underwater images.Experimental results on multiple benchmark underwater fish recognition datasets verify that this approach can effectively train large-scale fish recognition models to achieve a superior recognition accuracy that is approximately 5%higher than existing methods.(4)Adversarial learning based underwater fish recognition.Based on adversarial learning,this part of work proposes a simple,effective and generic robust model training method that can automatically distinguish the different importance of foreground an background to classification,and adaptively boost or suppress the learning of the fore-and back-ground.It thus can achieve more effective and robust model training with higher accuracies.It not only has theoretical novelty but also can effectively solve the learning bottleneck caused by noise background in underwater fish recognition task.The experimental results on three underwater image datasets show that the proposed adversarial training method Adv Fish can greatly reduce the dependency of the model on background noise and achieves 5% higher accuracy than existing methods on large-scale datasets.The research outcomes of this thesis not only provide a comprehensive experimental exploration with many key findings for the research field of underwater fish recognition,but also,through the three proposed novel machine learning methods,train a set of highperformance underwater fish recognition models.These models can be deployed in realworld scenarios to assist fishery plannings and marine ecological environment researches.The methods and experimental findings produced in this thesis are also informative for building general underwater vision models.They can also help promote the application of deep learning based computer vision in unconventional scenes.The three learning methods proposed in this thesis were verified on large-scale underwater image datasets,which are beneficial for training bigger models in the future.