End to End Segmentation of Canola Field Images Using Dilated U-Net

Abstract

Semantic segmentation is used in many fields like agriculture, medical imaging, and autonomous driving. The paper proposes an end to end solution for efficient weeds and crop segmentation in field environment application. The crop/weeds segmented output is utilized to generate a decision map for variable rate fertilizer and herbicide application. Currently available models are memory expensive and do not have real time performance unless enough computational power is accessible in field. We use Maximum Likelihood Classification (MLC) and image processing techniques to label field images in three classes; background, crop, and weeds. This data is processed through our modified U-Net, which improves the semantic accuracy with reduced memory cost. We train our model with DICE loss and compare the results with state of the art. We achieve 89.12% mean Intersection Over Union (mIOU) with 86.11%, 82.99%, and 98.23% individual IOU for crop, weeds, and background, respectively. Our proposed model uses only <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$15M$ </tex-math></inline-formula> parameters which are <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$57M$ </tex-math></inline-formula> less than the state-of-the-art models with a compromise of 1% mIOU score.