| Title |
A Matching-loss for Improving Deep Image Segmentation |
| Authors |
문구영(Goo-Young Moon) ; 김종옥(Jong-Ok Kim) |
| DOI |
https://doi.org/10.5573/ieie.2024.61.6.75 |
| Keywords |
Semantic segmentation; Small object detection; Mathing-loss; Crop images; Disease |
| Abstract |
Recently, in the filed of smart farm, there have been actively conducted studies to accurately detect disease regions by using segmentation for crop images. However, due to the small size of disease, predicting disease regions accurately is a challenging task. To address this issue, this paper proposes a novel optimization strategy, called Matching-Loss which stems from the translation invariance property of the segmentation network. Transformation invariance means that the feature embedding representation of the network for a transformed image should be identical to the transformed version of the feature embedding of the original input image. Matching loss offers the advantage of being easily calculated using Mean Squared Error (MSE). Furthermore, by leveraging the optimization based on the relationship between the input image and its transformed version, it is possible to enhance the generalization capability and robustness of the segmentation network. This improvement can lead to an enhanced recognition ability of the network for regions affected by various factors such as pests and diseases. We verified the significance of the matching loss through comparative experiments conducted on crop images captured in real field environments. |