University of Twente Student Theses


Image based bee health classification

Chawane, Shruti (2022) Image based bee health classification.

[img] PDF
Abstract:Honeybees are very important to the ecosystem and mankind as they play a vital role in the pollination process of the agricultural process and providing a balance to the bio-diversity of ecosystem. The state of the environment can also be deduced by keeping a track of bee health status. This leads to the need of a bee health classification system which can ubiquitously identify bee health status. Recent decades have seen research and development in the area of artificial intelligence to perform classification tasks from input images via a camera. Therefore, an image dataset with over five thousand images is taken into consideration for this project. This work presents the recent advances in the area in the last two to three decades, experiment with the state of the art models namely VGG-16 and DenseNet-121. The development methodology is inspired by the technique of transfer learning with pre-trained VGG-16 and DenseNet-121 on the ImageNet dataset. Initially, the reasons why the task of image based bee health classification differs from the pre-training of the deep models on ImageNet are mentioned. The reasons are also verified by the results of experiment 1. The results of experiment 1 provide a very clear picture as to what kind of features are extracted by both the models from images from ImageNet and the BeeImage dataset. The pre-trained models are observed to successfully extract fine and smooth features from underlying images that includes edges, texture, shape, surroundings of the objects. But the down stream task demands even finer features to detect bee health like orientation of bee, wing structure and quality, bodily deformities etc. In extension, the work establishes evaluation metrics for the task of image based bee health classification and critically analyses the results of model performances with evaluation metric of F1 scores and macro-F1 scores. The project also delivers on the best fine-tuning strategy for the image base bee health classification. In the end, it is found that DenseNet-121 based models outperform VGG-16 based models. Also, it is observed that both models have three 'feature-extracting blocks' along with two 'interpreting blocks'. It is deduced and concluded that the 'interpreting blocks' do indeed need fine-tuning to perform better than the "out of the box" pre-trained deep models. In the end, the research question derived with the project is answered conclusively.
Item Type:Essay (Master)
InsectSense, Wageningen, Netherlands
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Programme:Embedded Systems MSc (60331)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page