Wheat Plant Counting Using UAV Images Based on Semi-supervised Semantic Segmentation

Abstract

Plant counting in major grain crops like wheat through aerial images still poses a challenge due to the very high infield density of plants and occlusion. Annotation of aerial images for counting through perfect detection or segmentation is extremely difficult due to a large number of extremely small plant instances. In this paper, we present a semi-supervised method based on cross-consistency for the semantic segmentation of field images and an inception-based regression network for plant counting. Through loosely semantic segmentation, tiny plant clusters are extracted from the RGB image and fed to a regression network to get the count. Cross-consistency under the cluster assumption is a powerful semi-supervised training technique to leverage the unlabeled images. In this work, it is observed that regions with lower density are more detectable within hidden representations as compared to inputs. Supervised training of an encoder in a shared fashion and the main decoder is carried out on the RGB images and the corresponding mask. Consistency between the prediction of main and auxiliary decoders is imposed to leverage the unlabeled images. Induction of inception in the regression network benefits in extracting the multi-scale features which are very important because of quite tiny plant instances as compared to the whole image. The proposed plant counting framework achieves very high performance having a standard deviation of 0.94 and a mean of 0.87 of absolute difference in the count given the semi-supervised nature. Our network has performed reasonably well as compared to supervised detection and segmentation-based counting framework. Moreover, labeling for detection or segmentation is a quite tedious task, so our network has the leverage to train the model with few labeled and large numbers of unlabeled images which also provides the advantage to train the system for other crops like rice and maize with few labeled images.