Multi-Resolution Supervision Network with an Adaptive Weighted Loss for Desert Segmentation
نویسندگان
چکیده
Desert segmentation of remote sensing images is the basis analysis desert area. are usually characterized by large image size, large-scale change, and irregular location distribution surface objects. The multi-scale fusion method widely used in existing deep learning models to solve above problems. Based on idea feature extraction, this paper took results each scale as an independent optimization task proposed a multi-resolution supervision network (MrsSeg) further improve result. Due different difficulty branch task, we also auxiliary adaptive weighted loss function (AWL) automatically optimize training process. MrsSeg first lightweight backbone extract different-resolution features, then adopted module fuse local information global information, finally, multi-level decoder was aggregate merge features at levels get In method, treated AWL calculate adjust weight branch. By giving priority easy tasks, improved could effectively convergence speed model experimental showed that MrsSeg-AWL ability has faster speed, lower parameter complexity, more accurate results.
منابع مشابه
Multi-Resolution Language Grounding with Weak Supervision
Language is given meaning through its correspondence with a world representation. This correspondence can be at multiple levels of granularity or resolutions. In this paper, we introduce an approach to multi-resolution language grounding in the extremely challenging domain of professional soccer commentaries. We define and optimize a factored objective function that allows us to leverage discou...
متن کاملAn image segmentation method with multi-resolution mechanism
This report proposes a new image segmentation method with a multi-resolution mechanism. In this method, the high accuracy on image edge extraction and image texture evaluation is realized. The proposed method employs the edge extraction based on the average values of intensities in a small image part and also the texture evaluation based on the co-occurrence probability of a local image area. I...
متن کاملذخیره در منابع من
با ذخیره ی این منبع در منابع من، دسترسی به آن را برای استفاده های بعدی آسان تر کنید
ژورنال
عنوان ژورنال: Remote Sensing
سال: 2021
ISSN: ['2315-4632', '2315-4675']
DOI: https://doi.org/10.3390/rs13112054