Joint crop and tassel segmentation in the wild

Crop segmentation is a frequently concerned problem for computer vision applications in agriculture. Tassel is a typical agronomic trait in the crop breeding process. Tassel trait characterization also requires fine-grained shape extraction. However, previous methods are usually dependent of category, which is hard to transfer to other cultivars with different colors. To address this, the goal of this study is to develop a feasible method that can deal with different categories simultaneously and that is easy to transfer. Targeted on maize, we proposed to jointly segment crop and maize tassel. The task is consequently formulated as a semantic segmentation problem. We proposed a region-based approach that leverages the efficient graph-based segmentation algorithm and simple linear iterative clustering (SLIC) to generate region proposals. Then, a neural network based color model is learnt to execute the semantic labeling. We demonstrate the effectiveness of our method on two typical crop and tassel dataset respectively. Experimental Results show that our approach significantly outperforms other state-of-the-art approaches on the tassel segmentation and achieves comparable performance on the traditional crop segmentation. Results of this research can serve to the agriculture automation, mechanization and intellectualization. The dataset and source code are made available online.