An advanced Bangladeshi local fish classification system based on the combination of deep learning and the internet of things (IoT)

Abstract

Fish classification leads to the automated machine-based fish separation system. In terms of classification and real-time data monitoring, deep learning and the Internet of Things (IoT) each provides an efficient solution. This paper focuses on the development of an embedded system based on the principles of Deep Learning and IoT. The proposed methodology is classified into interconnected parts. The first part describes the working principles of DL with along the dataset building, model analysis and overall system architecture. A new dataset from eight different Bangladeshi fish species. In the process of DL, First, two sets of datasets have been created namely, setup-1(S1) containing original images and setup-2(S2) containing Unsharp masked photos. Then, seven conventional ImageNet pertained state-of-the-art deep learning models on both benchmarking setups: InceptionV3, Xception, DenseNet121, DenseNet169, DenseNet201, InceptionResNetV2, and ResNet152V2. In the process of IoT, the architectural design of a smart contained has been deployed with the aid of several kinds of sensors and microcontrollers. This research has found satisfactory results with the DL models and IoT-based components. The best benchmark accuracy for setup-1 was 96% for all of the DenseNet121, DenseNet169, and DenseNet201 architecture, and for setup-2, it was 96% for the Xception model. Finally, we have constructed a hybrid (CNN + Convolutional LSTM) model, for which the accuracy was 97%, outperforming all of the abovementioned state-of-the-art methods. Besides, the research has performed some experiments with the IoT-based Solution. Though the proposed solution has exhibited some drawbacks, but it can be practicable in real-time solutions.