Performance evaluation on quantized weight for convolutional neural network based object detection
A Convolutional Neural Network (CNN) based object detection is an emerging topic in the image processing field and has become the state-of-the-art in computer vision and machine learning. The traditional system in object detection uses a handcrafted feature extractor which is less robust in applicat...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English English |
| Published: |
2021
|
| Subjects: | |
| Online Access: | http://eprints.utem.edu.my/id/eprint/26088/1/Performance%20evaluation%20on%20quantized%20weight%20for%20convolutional%20neural%20network%20based%20object%20detection.pdf http://eprints.utem.edu.my/id/eprint/26088/2/Performance%20evaluation%20on%20quantized%20weight%20for%20convolutional%20neural%20network%20based%20object%20detection.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | A Convolutional Neural Network (CNN) based object detection is an emerging topic in the image processing field and has become the state-of-the-art in computer vision and machine learning. The traditional system in object detection uses a handcrafted feature extractor which is less robust in applications. By applying the CNN approach in the field, the accuracy of object detections can increase significantly. However, the use of deep CNN architecture model leads to high computation. In this research, a real-time CNN based object detection system is presented. The system is designed based on the modified You Only Look Once (modified-YOLO) architecture which is constructed with only 7 CNN layers. The grid cell parameter value of the system is varied to evaluate its effectiveness and ability in detecting small size objects upon deployment. The experimental results demonstrate that even with 7 convolutional layers, modified-YOLO can provide good detection accuracy and real-time operation achieving the best miss rate (MR) of 22.7% MR. Although the scores show an increase in the MR, the visual qualitative evaluation using randomly captured images indicate that the 7 layers modified-YOLO architecture with 11x11 grid cells can correctly and easily detect small objects. This makes the modified YOLO architecture which has been reduced in terms of complexity a suitable candidate for use in real-time operation. In order to further reduce the complexity of the CNN system, the trained floating-point weights are quantized. Three types of scalar quantization are used to quantize the CNN weights namely symmetric uniform quantizer, asymmetric uniform quantizer and non-uniform quantizer designed using k-means algorithm. The quantization reduces the storage and computation requirements. The quantitative results using the MR standard metric indicate that the non-uniform quantizer provides the best results compared to the other quantization methods. Using 6-bit precision non-uniformly quantized weights yields detection performance comparable to the CNN network using floating-point weights. Additionally, based on the qualitative results, the CNN network with 4-bit non-uniform quantization weights is able to detect the person objects correctly. |
|---|