Year: 2024
Type: article
Abstract: This work investigates the enhanced capabilities of transfer learning in computer vision, specifically focusing on object detection. We employ Dense Convolutional Network (DenseNet) architectures, renowned for their efficacy in extracting and reusing features for the purpose of object identification in video frames. These advanced networks can circumvent the generally substantial requirements for computational resources and time-consuming training associated with deep learning tasks by utilizing transfer learning techniques. Our research aims to optimize the performance of DenseNet121, DenseNet169, and DenseNet201 models by making minor adjustments using a meticulously selected dataset. The objective is to assess their efficacy in identifying and categorizing items. The results indicate that DenseNet121 and DenseNet201 attain remarkable validation accuracies of 0.9605, while DenseNet169 closely matches with an accuracy of 0.9585. Results showcase the versatility of DenseNet models in tackling various object detection tasks and their comparable performance levels. This means that any of the three models might work for real-world picture recognition tasks. Moreover, our discoveries establish a basis for future investigations into enhancing the reliability of models for real-world use in monitoring and autonomous car systems, capitalizing on the models' proven accuracy.
Authors: Sara Bouraya, Abdessamad Belangour
Institution: University of Hassan II Casablanca
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W4404006334