Click on “Download PDF” for the PDF version or on the title for the HTML version.

If you are not an ASABE member or if your employer has not arranged for access to the full-text, Click here for options.

Highlights

We have established an image dataset named GrainPest, specifically designed for detecting stored-grain pests on the surface of grain bulks.

Abstract. Pest detection plays an important role in integrated pest management (IPM). An effective dataset is the foundation for achieving the automatic detection of stored-grain pests. Although some datasets of stored-grain pests have been published, they are not suitable for detecting pests on the surface of grain bulks because the image backgrounds and the shooting process have significant differences. To provide data for the development of algorithms for pest detection on the surface of grain bulks in granaries, we established an image dataset named GrainPest in this study. It contains 16358 images, including six categories of pest instances. A total of 66372 pest instances with detection bounding boxes were annotated. GrainPest exhibits inter- and intra-class variance, data imbalance, and small-size pest instances, most of which occupy less than 0.25% of the total pixel area of the entire image. The hard negative sample images containing pest-like objects that easily cause misdetection, such as impurities, mold spots, shadows between different grain kernels, and blackened corn ear bases, were collected. For small pest detection, we proposed that improving the performance of the algorithms by simply increasing the size of the feature map can be effective, which was demonstrated in subsequent experiments. We also demonstrated that incorporating hard negative sample images into algorithm training can mitigate the issue of mis-detection. We conducted baseline experiments using object detection algorithms (i.e., YOLOv5, Faster R-CNN, CenterNet, and DETR) and image classification algorithms (i.e., VGGNet and ResNet). Experimental results show that GrainPest can provide effective data support for the research of pest detection on the surface of grain bulks, small object detection, and image classification. We have made GrainPest publicly available at www.grainpest.com