Low-Cost Domain Adaptation for Crop and Weed Segmentation

Abstract

The use of new and more sustainable technologies in agriculture is important to reduce the need for agrochemicals and improve energy efficiency. Many of the recent approaches in this area are based on computer vision algorithms. However, due to the great variability of the scenes that can occur in an agriculture field, the domain shift phenomenon is a relevant problem in this area. Domain adaptation attempts to mitigate this data variability problem. In this work, we propose a low-cost domain adaptation method between agriculture domains for image segmentation. Our approach performs domain adaptation by changing the amplitude of the low-frequency spectrum of images along with contrast limited adaptive histogram equalization (CLAHE), which is an efficient replacement for a commonly used image-to-image translation methods, such as Cyclegan, drastically reducing the number of parameters and improving quality in cases where these models do not maintain semantic consistence between translations.