Performance Measurement of UNet and DeepLabV3Plus for Semantic Segmentation of Weeds and Crops.

Sammanfattning

The objective of semantic segmentation is to predict classes for all pixels in an image. In the context of agriculture, there is a need for models that predict where crops and weeds are in an image of a plantation. This would allow automatic removal of weeds. In this work, a UNet was built and compared with DeepLabV3+, in order to see which combination of models and ResNet backbones performed best after a day of training. The results showed that DeepLabV3+ performed considerably better than UNet when trained and tested on two subsets of the SugarBeets dataset. A few augmentations were used in two ablation studies, which showed that augmentations had a worse effect on the performance when trained on dataset 2. A similar conclusion was made when the models were trained with augmentations using dataset 1, except that UNet performed better than it did without augmentations.