Development of morphometric and mathematical transform based systems for classifying fishes using machine learning techniques

Abstract

Fish is an important resource for humans in terms of both health and commerce. Since newlinefish meat has excellent taste and is a good source of protein and other nutrients, many newlinefish species are consumed as food around the world. India s fish and fish products make newlineup more than 10% of the country s total exports of agricultural goods, with over one newlinemillion tonnes exported. Export of various ornamental and commercial food fish species newlinecontributes largely to the Indian economy. To export raw fish and processed fish products, newlinefish must be sorted into species types in the fish trade sector. Separating fishes manually isa time-consuming and error-prone process. Evolution in image processing and machine newlinelearning techniques led to the development of automated systems for separating fish newlinespecies. newlineThe thesis presents several models for the classification of fish species using advanced newlinetechniques in image processing and machine learning. From an industrial point of view, newlinethe study focuses on the classification of Wrasse fishes in the ornamental fish category and newlinetuna fish in the commercial food fish category. Wrasse fishes are classified into genus and species levels using a multidomain feature-based system and an extreme learning machine (ELM) classifier. Using feature descriptors and pre-trained convolutional neural network newline(CNN) based models, commercial food fishes are initially divided into economically newlinerelevant tuna species and other fishes. Further, the classification of commercially relevant newlinetuna into species types is evaluated using a conventional transform-based approach and newlineusing CNN architectures. newlineMultidomain features are employed in the categorization of ornamental Wrasse fishes. newlineColor histogram, local binary patterns (LBP), histogram of oriented gradients (HOG), newlineand statistical characteristics from wavelet transform sub bands are used in this approach. newlineFor feature reduction, an ensemble of dimension reduction algorithms is employed. The performance of the combined and reduced feature sets for both genus and species