| Author | Natthasit Wongsirikul |
| Note | A thesis submitted in partial fulfillment of the requirements for the
degree of Master of Sciences in
Microelectronics and Embedded Systems |
| Publisher | Asian Institute of Technology |
| Abstract | Deep learning such as Convolution-Neural-Network (CNN) has taken over the field of
computer vision when it comes to the application of object detection. The popularity of CNN is due in part to its superior performance over other traditional image processing techniques. In addition, CNN-based models such as RCNN and the Yolo allows transfer learning where the models can be trained to detect specific objects by utilizing the already
robust feature extraction which are trained by a massive datasets such as the PASCAL
VOC. This allows the models to achieve high performance even though it is trained on
smaller dataset. For these reasons, CNN-based models have become the top choice for target-specific object detection applications. However, these models were designed to run on desktop environment, often with GPU support. They are not optimized to run on an embedded system, which has lower computation power and memory space. For a trained CNN-model to be inference ready, some optimization must be performed to make it implementable out in the field. In this thesis, some popular CNN-based model optimization techniques are explored. A compression algorithm is developed based on a method called filter pruning. The compressed models are then compiled to run on an embedded system where their performance, speed, and memory usage were examined
against its non-compressed counterpart.
|
| Year | 2019 |
| Type | Thesis |
| School | School of Engineering and Technology (SET) |
| Department | Department of Industrial Systems Engineering (DISE) |
| Academic Program/FoS | Microelectronics (ME) |
| Chairperson(s) | Mongkol Ekpanyapong; |
| Examination Committee(s) | Dailey, Matthew N. ;Abeykoon, A.M. Harsha S. ; |
| Scholarship Donor(s) | Royal Thai Government Fellowship; |
| Degree | Thesis (M. Sc.) -- Asian Institute of Technology, 2019 |
