Design of an efficient multi-objective recognition approach for 8-ball billiards vision system
In this paper, some key technologies based on colour image processing for 8-ball billiards robot vision system are discussed and an efficient approach for multi-objective recognition is proposed. This approach is divided into two parts, i.e. multi-objective detection and ball pattern recognition. I...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2018-01-01
|
| Series: | Kuwait Journal of Science |
| Subjects: | |
| Online Access: | https://journalskuwait.org/kjs/index.php/KJS/article/view/2083 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | In this paper, some key technologies based on colour image processing for 8-ball billiards robot vision system are discussed and an efficient approach for multi-objective recognition is proposed. This approach is divided into two parts, i.e. multi-objective detection and ball pattern recognition. In image pre-processing, the normalized RGB colour space and histogram statistics are adopted for segmentation of background (table cover) and foregrounds. In order to accurately locate and isolate the single ball in a local region, the improved Hough Transform (HT) algorithm and the Least Squares (LS) method are adopted in combination. The improved HT algorithm is used for the purpose of eliminating the noise concentrated at edge points, and the LS method is used for fitting the circle center accurately with the least mean square error. Based on single ball detection in a local region, the multi-ball detection approach has been worked out to locate the position of each ball on the table. In the experiment, the proposed approach has been proved to complete the detection with an accuracy of 99.4% in 0.65s in average, and the performance is better
than the traditional Circular Hough Transform (CHT) algorithm and the K-means cluster method. In addition, the Convolution Neural Network (CNN) method is adopted for pattern recognition of each target ball being segmented, i.e. identification of a solid ball or a striped ball. In order to improve the quality of CNN training: the colour segmentation and morphologic operation are applied for the segmented ball image pre-processing; the training set images are rotated for augmentation; pre-training stage is introduced in for optimizing the initial weight matrices. The calibrated image blocks are imported to the network for training. In the verification test, the trained CNN model shows a recognition rate of over 98.5%, and outperforms the other three classic methods. The introduction of CNN method has been proved to be correct and effective, and is an innovative and significant step
for the design process of the 8-ball billiards robot vision system.
|
|---|---|
| ISSN: | 2307-4108 2307-4116 |