A new method is proposed to achieve large-scale cigarette filters detection under a wide field of view. It is expected to obtain accurate detection under the situation of overlapping, sheltering, distortion and low contrast. Firstly, by adding the spatially aware based the self-attention argument module and a Focal Loss function, an improved U-net model named as SAAU-net is proposed, which can achieve highly accurate semantic segmentation. After receiving a correct segmentation mask, the circle center of the cigarette filters is determined by object detection. Based on the circle theorem, the structural element matching is used to detect the circle centers. Hidden Markov model (HMM) is employed for direction searching. Experiments performed in simulation and industry application environments (with 5000 boxes) verify that the accuracy of the proposed approach can achieve 99.95%. Results also show that the proposed method has strong robustness to adapt different challenging environments.
|Translated title of the contribution||Investigation of accurate detection of large-scale objects under a wide field of view|
|Original language||Chinese (Traditional)|
|Number of pages||9|
|Journal||Yi Qi Yi Biao Xue Bao/Chinese Journal of Scientific Instrument|
|Publication status||Published - 01 Apr 2020|
Bibliographical notePublisher Copyright:
© 2020, Science Press. All right reserved.
Copyright 2020 Elsevier B.V., All rights reserved.
- Focal loss
- Hidden Markov model
- Large scale objects
- Self-attention argument
- Wide field of view
ASJC Scopus subject areas