Abstract—We develop a monitoring and forewarning system to detect planthoppers in paddy fields. Our detection algorithm consists of two stages. At the first stage, we extract the main paddy in the middle of an image by some traditional image processing techniques. At the second stage, we use a convolutional neural network to detect planthoppers within the extracted region. Our detection model is revised from the Single Shot MultiBox Detector (SSD). The original SSD model usually misrecognizes reflected light as planthoppers since a lot of background information has been discarded in the max pooling layers of the SSD model. To solve this misrecognition problem, we propose a new kind of pooling-- Local Difference Pooling. This proposed method greatly improves the performance of planthopper detection to achieve 89.38% precision and 91.93% recall.
 

Index Terms—planthopper, detection, SSD, local difference pooling