U-SegNet with Parallel Pooling Attention for Crop Pest Detection

Abstract Crop pests seriously affect the yield and quality of crops. Early pest detection is the premise and basis of scientific pest control in modern intelligent agriculture. However, due to the small size, and irregular shapes, postures, colors and appearances of pests in the field, the detection of pests is always a challenging problem. To solve the problems of the traditional pest detection methods, such as low identification accuracy and vulnerability to orientation, shape, illumination and complex background, a modified U-SegNet model with parallel pooling attention (U-SegNet-PPA) is proposed for pest detection. It is a symmetric hybrid architecture consisting of U-Net, SegNet and parallel pooling attention (PPA), where skip connection and incorporates fine multi-scale feature are used to describe the pixels and boundary of the pest images, and PPA is added into U-SegNet to capture the compositional semantic features of the pest image. Comparing with fully convolutional network (FCN), U-Net, SegNet and U-SegNet, it has some advantages of fast training speed and few model parameters, and attention mechanism and data augmentation are utilized to greatly improve the performance of pest detection in field. The experimental results on our newly collected crop pest image dataset show that the proposed approach performs better than other state-of-the-art methods. It can provide a valuable support for the detection, identification and severity estimation of crop pests.