Dandelion segmentation with background transfer learning and RGB-attention module

Abstract

Automatic weed control using weeding robots has arisen as a promising alternative for reducing the amount of herbicide applied in the agriculture and landscaping industry. Dandelion detection is a challenging task for weeding robots. To obtain accurate dandelion segmentation results, a multi-category segmentation dataset is converted into a binary segmentation dataset to learn the difference between background and foreground, which is referred to as background transfer learning. This method can help us use different datasets to overcome the problem of dandelion segmentation in a complex background environment. In addition, the color space information of the dandelion images with concentrated colors is aggregated in convolutional networks. Inspired by the attention mechanism in the neural network, we design a color-attention module to learn the weight of each color space. In this way, the color information is used to segment the dandelion. By combining background transfer learning and color-attention module methods, the dandelion segmentation method can effectively segment the dandelion with a satisfactory accuracy rate. The results of ablation experiments show that for the input images of different colors, the method proposed in this paper can learn different attention information. The proposed method can be used in agricultural and landscape industry weeding.