Dilated inception U-Net with attention for crop pest image segmentation in real-field environment

Abstract

Automatic pest image segmentation (PIS) plays a vital role in pest detection and recognition. However, it remains a difficult issue due to the various irregular pest images and low contrast between pests and their surroundings. A dilated Inception U-Net with attention (DIAU-Net) is constructed for PIS. It is a U-shape encoder–decoder multi-scale convolution model, consisting dilated residual Inception (DRI), multi-scale feature fusion (MSFF), and multi-scale dilated attention (MSDA), where DRI instead of the convolution is employed to capture the multi-scale local features, MSFF is added into the bottleneck layer to extract the semantic information, and MSDA instead of skip connection is used to fuse the extracted low-level features and high-level features. Experimental results on a crop pest image dataset validate that DIAU-Net based PIS method outperforms other state-of-the-art PIS methods, with Dice score of 93.12 % compared to 82.35 % for the U-Net based method. The proposed method can provide valuable support for the detection, identification and severity estimation of crop pests in real field environment.