MSc. CIS Theses and Dissertations (2019)

Permanent URI for this collection

http://hdl.handle.net/11071/6603

Browse

Browsing MSc. CIS Theses and Dissertations (2019) by Subject "Fine-grained classification"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    A Mobile-based image recognition system for identifying bird species in Kenya
    (Strathmore University, 2019) Nyaga, Gideon Mwangi
    Kenya is world-renown for its wildlife tourism, which attracts many foreigners and is a leading foreign exchange earner. The last few years has seen significant growth in the number of people, called birders, who observe birds for recreation or for citizen science. Kenya in particular is one of the leading birding destinations in the World with about 11% of the World’s bird species. A big challenge birders face is correctly identifying the bird species observed. Fine-grained visual classification, and in this case the ability to identify Kenyan bird species, is a challenging task for both humans and machines mainly because of subtle differences between different bird species and strong variety within same species. This research study solved this challenge by developing a mobile-based image recognition system that can identify Kenyan bird species from images. A deep neural network called a Convolutional Neural Network (CNN), inspired by the human visual cortex, is well suited for image recognition because it is able to handle shifts and distortions in an image well, has fewer trainable parameters and thus uses less memory and training time than a standard artificial neural network (ANN). Further a depth-wise separable CNN is suited for resource-constrained mobile devices as it has lower computational cost as compared to standard CNNs. This research developed a depth-wise separable CNN model, which was trained using 39,031 labelled bird images via supervised transfer learning. The model achieved a final test accuracy of 97.3%. Consequently, an Android mobile application was developed to consume the resulting model. The model was embedded into the mobile application. Therefore, the user did not need internet to make an inference. The mobile application was able to process an input image containing a bird and identify the bird species. The mobile application was also able to give more details of the identified bird species.