Instance Segmentation Guided by Weight Map with Application to Tooth Boundary Detection
0
조회
0
다운로드
2
인용
0
참고문헌
키워드
초록
In this paper, we propose a method based on a weight map to improve the performance of instance segmentation and demonstrate the method using a simple application. A weight map is a set of pixel-wise losses, each of which has a different value depending on whether the pixel is located on the border of the image. Importantly, the losses of pixels on the border have a relatively higher value than the other pixels so that they can impose heavy penalties in the training stage. We verified the effectiveness of our method by assessing its performance when processing clinical dental images. Because teeth have similar image features (e.g., color, shape), and as they are arranged side by side, it is appropriate to evaluate the effect of using a weight map. With reference to the weight map, Mask R-CNN, our baseline model, learns very small, narrow boundaries to distinguish different instances that were recognized as one instance before. The improvement is evident both quantitatively and qualitatively. The Average Precision and Recall were found to have increased by 4.4% and 7.3%, respectively, with weight map learning. Thus, the proposed method was demonstrated to effectively enhance the ability to detect subtle boundaries. This finding is expected to make it possible to utilize a variety of existing models to their fullest potential.
Abstract
In this paper, we propose a method based on a weight map to improve the performance of instance segmentation and demonstrate the method using a simple application. A weight map is a set of pixel-wise losses, each of which has a different value depending on whether the pixel is located on the border of the image. Importantly, the losses of pixels on the border have a relatively higher value than the other pixels so that they can impose heavy penalties in the training stage. We verified the effectiveness of our method by assessing its performance when processing clinical dental images. Because teeth have similar image features (e.g., color, shape), and as they are arranged side by side, it is appropriate to evaluate the effect of using a weight map. With reference to the weight map, Mask R-CNN, our baseline model, learns very small, narrow boundaries to distinguish different instances that were recognized as one instance before. The improvement is evident both quantitatively and qualitatively. The Average Precision and Recall were found to have increased by 4.4% and 7.3%, respectively, with weight map learning. Thus, the proposed method was demonstrated to effectively enhance the ability to detect subtle boundaries. This finding is expected to make it possible to utilize a variety of existing models to their fullest potential.
참고문헌 (22)
- Kaiming H, "Mask r-cnn", Ieee I Conf Comp Vis, pp.2961~2969 (2017)
- Zhaowei C, "Cascade r-cnn: high quality object detection and instance segmentation", Ieee T Pattern Anal (2019)
- Jialian W, "Forest r-cnn: large-vocabulary long-tailed object detection and instance segmentation", ArXiv Preprint ArXiv: 2008:05676, ~ (2020)
- Yiqing Z, "Mask-refined r-cnn: a network for refining object details in instance segmentation", Sensors, 20(4), pp.1010~ (2020)
- Shaoqing R, "Faster r-cnn : towards real-time object detection with region proposal networks", Adv Neur In, pp.91~99 (2015)
- Jonathan L, "Fully convolutional networks for semantic segmentation", Proc Cvpr Ieee, pp.3431~3440 (2015)
- Antoni, Bartomeu C, "A non-local algorithm for image denoising", Proc Cvpr Ieee, 2, pp.60~65 (2005)
- Xiaolong W, "Non-local neural networks", Proc Cvpr Ieee, pp.7794~7803 (2018)
- Yue C, "Gcnet: non-local networks meet squeeze-excitation networks and beyond", Ieee I Conf Comp Vis Workshops (2019)
- Kai C, "Hybrid task cascade for instance segmentation", Proc Cvpr Ieee, pp.4974~4983 (2019)
- Xinyu L, "Global context parallel attention for anchor-free instance segmentation in remote sensing images", Ieee Geosci Remote S (2020)
- Di L, "Zigzagnet: fusing top-down and bottom-up context for object segmentation", Proc Cvpr Ieee, pp.7490~7499 (2019)
- Shu L, "Path aggregation network for instance segmentation", Proc Cvpr Ieee, pp.8759~8768 (2018)
- Yongliang Q, "Cattle segmentation and contour extraction based on mask r-cnn for precision livestock farming", Comput Electron Agr, 165, pp.104958~ (2019)
- Chuang Y, "Segmentation and measurement scheme for fish morphological features based on mask r-cnn", Information Processing in Agriculture (2020)
- Karen S, "Very deep convolutional networks for largescale image recognition", ArXiv Preprint ArXiv: 1409:1556, ~ (2017)
- Matthew DZ, "Visualizing and understanding convolutional networks", European Conf. on Computer Vision, pp.818~833 (2014)
- Kaiming H, "Deep residual learning for image recognition", Proc Cvpr Ieee, pp.770~778 (2016)
- Jan H, "Learning non-maxium suppression", Proc Cvpr Ieee, pp.4507~4515 (2017)
- Olaf R, "U-net: Convolutional networks for biomedical image segmentation", Int’l Conf. on Lect Notes Comput Sc, pp.234~241 (2015)
- "LabelMe: image polygonal annotation with python" (2020)[URL]
- 민진홍, "Grammatical Error Correction Models for Korean Language via Pre-trained Denoising", Quantitative Bio-Science, 39(1), pp.17~24 (2020), 자연과학연구소[DOI]