EW-Fisher: A Novel Loss Function for Deep Learning-Based Image Co-Segmentation
- PDF / 1,269,439 Bytes
- 15 Pages / 439.37 x 666.142 pts Page_size
- 50 Downloads / 162 Views
EW-Fisher: A Novel Loss Function for Deep Learning-Based Image Co-Segmentation Xiaopeng Gong1
· Xiabi Liu1 · Xin Duan1 · Yushuo Li1
Accepted: 15 September 2020 © Springer Science+Business Media, LLC, part of Springer Nature 2020
Abstract The loss function is an important factor for the success of machine learning. This paper proposes a new loss function for deep learning-based image co-segmentation. It aims to maximize the inter-class difference between the foreground and the background and at the same time minimize the two intra-class variances. This idea has some similarity to the Fisher criterion in pattern recognition. We further embed an edge weighting strategy into this form of Fisher-like criterion to let the pixels near the foreground edges be paid more attentions in the training process for achieving the finer segmentation. The resultant loss function is called EW-Fisher (Edge-Weighted Fisher). We apply the proposed EW-Fisher loss to image co-segmentation and evaluate it on commonly used datasets. In the experiments, the EWFisher stably outperforms the most-widely used cross-entropy loss and Dice loss as well as the recently presented edge agreement loss and Hausdorff distance loss. The comparison results and the ablation studies prove the values of our Fisher-like learning criterion and edge weighting strategy. Keywords Loss function · Deep learning · Image co-segmentation · Edge weighting strategy · Fisher criterion
1 Introduction In recent years, the deep learning has attracted a lot of attentions and has gained encouraging results in solving the problem of image segmentation. A critical factor for the success of such
B
Xiabi Liu [email protected] Xiaopeng Gong [email protected] Xin Duan [email protected] Yushuo Li [email protected]
1
Beijing Lab of Intelligent Information Technology, School of Computer Science, Beijing Institute of Technology, Beijing 100081, China
123
X. Gong et al.
an approach is the loss function. Currently, the cross-entropy (CE) [1] and the Dice [2] are two most widely used loss functions in the field of deep learning-based image segmentation. The CE loss measures the quality of image segmentation from the view of single pixels. It aims at the correct classification of each pixel in the image [3–5] and ignores the integrality of foreground or background objects. So, if most of pixels have been classified correctly, the segmentation quality is not easy to be improved further. More importantly, the two classes are usually highly unbalanced in image segmentation. For example, the interested foreground objects often occupy only a small part of the image. The CE loss is unsatisfactory to deal with the such class unbalance problem. The focal loss was presented to alleviate this shortcoming of the CE loss by assigning different balance coefficients to different classes, in order that the training will not be overwhelmed by some classes with much more instances [5]. However, it still focuses on the classification of single pixels and ignore the integrity of foreground or
Data Loading...
