Abstract:
Ducks in Bangladesh are an integral part of the country’s biodiversity. Ducks play
significant role in both ecological balance and local livelihood. Duck meat is a rich source
of protein. There are somewhere 6-7 duck species available in Bangladesh. Each species
holds different characteristics. Duck species classification includes the implementation of
transfer learning and image processing techniques which leverages some pre-trained
models, ResNet50, ResNet152, DenseNet121, DenseNet201 and MobileNet-v2. Accurate
classification of duck species like Campbell Female, Campbell Male, Deshi Female, Deshi
Male, Runner Female, Runner Male, Zinding Female and Zinding Male is the main
motivation of this study. The methodology includes proper data collection from a hatchery
in Bikrampur and been proceed through some pre-processing methods for ensuring the data
quality. The performance of the implemented model followings: ResNet50 achieved
75.00% of accuracy with test data loss of 0.795, ResNet152 achieved 60.50 % of accuracy
with test data loss of 1.043. DenseNet121 achieved 98.25% and test loss is 0.081, and
MobileNet-v2 achieved 97.75% accuracy with 0.070 data loss. The proposed model,
DenseNet201 abled to achieve 99.25% accuracy with test loss of 0.027. However, many
researchers had some limitations of limited classes, real-time images and other aspects after
model training. The study impacts on society, environment by providing ecosystem
balance, enhanced monitoring techniques and better identification of the ducks. System’s
potential is also been considered by emphasizing ethical aspects which recommends some
guidelines for responsible use of technology. The research participates on development of
a precise and automated system of duck species classification. Future study will focus into
enhanced model’s decision making, exploring real-life monitoring for continuous health
monitoring and lifestyle of the ducks of various species.
