Institutional Repository of Vilnius University
Cassava leaf disease recognition using computer vision /
Abstract
dc:descriptionAbstract This paper explores the development and refinement of innovative deep-learning ar- chitectures for image classification tasks of classifying cassava leaf diseases. The backbone of this research is the EfficientNetB0 model, one of the most widely-utilized Convolutional Neural Networks developed by Google. Various methods to address class imbalance issues were explored, including the use of sigmoid focal loss, class weights, and over/under-sampling techniques. The most promising results were observed in the over/under-sampling approach. To enhance per- formance, the number of MBConv blocks within EfficientNetB0 was reduced, leading to improved classification outcomes and a lighter model, requiring less memory and resources to operate. The SE attention module was replaced with the ECA module, successfully reducing the model size and boosting its accuracy, thereby advancing the model’s efficiency. Hybrid architectures were also investigated, with a transformer encoder connected to the EfficientNetB0 output, where 7x7 feature maps were utilized for classification. Although a slight decrease in accuracy was recorded with the cassava leaf disease dataset, a performance improvement was observed when applied to the AID dataset.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2023
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Petrikas, Aivaras,
Subjects
dc:subject × 1Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- eng
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD192847160&prefLang=en_US