An Attention Mechanism Approach for Natural Color Retrieval for Plant Monitoring

Abstract

Even though agriculture practices have been continuously developed with the support of modern technologies, many more improvements can be made to enhance agricultural technologies and businesses. One such technology is the use of specific light color combinations to optimize the growth rate of plants. One obvious drawback is that plants’ color will change according to the light color combinations. The light color can fool human eyes and may cause errors when monitoring for plant anomalies. Color correction methods should be applied to help restore the natural plant color with the white light source from the unnaturally colored plant images. Our color correction method uses an application of self-dot-product attention, multi-head attention, and channel attention combined with a U-Net-based model. This proposed method performs the color correction with the input image in the RGB color space in two steps. First, a global transformation network provides the global function that maps the input RGB color vectors from every pixel and produces the corrected RGB color vectors. The global mapping function is the same for all pixels in the image. Next, a local transformation network attempts to correct the local color distortions such as light the flickering of LED light due to the AC power supplier.