Vol.11 (2021) No. 3 ISSN: 2088-5334 Local Freshwater Fish Recognition Using Different CNN Architectures with Transfer Learning Anup Majumder a,* , Aditya Rajbongshi a , Md. Mahbubur Rahman b , Al Amin Biswas c a Department of CSE, Jahangirnagar University, Savar, Dhaka, 1342, Bangladesh b Software Engineer, Crowd Realty, Tokyo, 1700005, Japan c Department of CSE, Daffodil International University, Dhanmondi, Dhaka, 1207, Bangladesh Corresponding author: * anupmajumder@juniv.edu Abstract— Bangladesh has its profusion of water resources, but due to environmental issues and some other significant causes, the quantity of water resources is lessening continuously. As a result, many of our local freshwater fishes are being abolished, leading to a lack of knowledge about freshwater fish among the new generation of people in Bangladesh. It is also very difficult to recognize freshwater fish because of their nature, color, shape, and structure. To recognize the local freshwater fish efficiently, transfer learning can be used, one of the significant parts of deep learning that concentrates on storing knowledge gained while solving one problem and employing it to a distinct but related problem. This paper has used six CNN-based architecture with transfer learning, namely DenseNet201, InceptionResnetV2, InceptionV3, ResNet50, ResNet152V2, and Xception. A total of seven freshwater fish image data is used here, which is collected from the various local fish markets of Bangladesh. To check the effectiveness of the working approach, we have calculated the accuracy, precision, Recall, and F1-Score. The approach InceptionResnetV2 and Xception achieved the highest accuracy with 98.81% over the other approach which is a very significant result. Keywords— Freshwater local fish; recognition; transfer learning; CNN. Manuscript received 30 Dec. 2020; revised 9 Feb. 2021; accepted 16 Feb. 2021. Date of publication 30 Jun. 2021. IJASEIT is licensed under a Creative Commons Attribution-Share Alike 4.0 International License. I. INTRODUCTION Bangladesh is a riverine country with bunches of waterways jumbled from East-West-North-South. The outpouring of water from Bangladesh is the third most elevated on the planet, close to the Amazon and Congo frameworks. From 2000 and 2016, aquaculture production increased from 712,640 and 2,060,408 metric t, a much larger quantity than wild capture production (1.023 million t) in 2016. Total fish production in Bangladesh in 2014–2015 was reported to be 3,684,245 MT, of which 1,023,991 MT (27.79%) were from inland open waters, 2,060,408 MT (55.93%) from inland closed waters and 599,846 MT (16.28%) from marine fisheries [1]. Almost 795 native species of fish and shrimp species of fishes cultivate in Bangladesh [2]. Everyday people must have some fish on their food plate. Many of the freshwater fishes of Bangladesh are being abolished for several reasons, which leads to a lack of knowledge about freshwater fish among the people in Bangladesh. Recently a lot of work has been done to recognize fishes though fish recognition is mainly restricted to constrained environments. As fish recognition is a challenging issue, researchers continuously applied several novel approaches to find the best results. Alsmadi et al. [3] proposed a methodology to recognize fishes using neural networks. In this work features are extracted by combining the size and shape of the fish images. The system has been done on distinct 350 fish images of 20 fish families and in the overall work acquired 86% accuracy applying the neural network incorporated with the back- propagation algorithm. Rahman et al. [4] proposed a methodology to recognize the Local birds of Bangladesh using MobileNet and Inception-V3. They used 7 types of local birds and augmented the imagedata using image processing techniques. The highest accuracy is found for MobileNet model with transfer learning technique. A Computer vision-based approach and neural network is applied to recognize fish species by Storbeck and Daan [5]. This method achieved the accuracy of 98% but they did not mention anything about the dataset. Montalbo and Hernandez [6] proposed a methodology to recognize fish species using VGG16 Deep Convolutional Neural Network (DCNN). Though this approach gets 98.67% accuracy, they have used 1078