Improving the generalization capability of YOLOv5 on remote sensed insect trap images with data augmentation

Abstract

AbstractIn agricultural pest management, the traditional insect population tracking in the case of several insect types is based on outsourced sticky paper traps that are checked periodically by a human operator. However, with the aid of the Internet of Things technology and machine learning, this type of manual monitoring can be automated. Even though great progress has been made in the field of insect pest detector models, the lack of sufficient amount of remote sensed trap images prevents their practical application. Beyond the lack of sufficient data, another issue is the large discrepancy between manually taken and remote sensed trap images (different illumination, quality, background, etc.). In order to improve those problems, this paper proposes three previously unused data augmentation approaches (gamma correction, bilateral filtering, and bit-plate slicing) which artificially enrich the training data and through this increase the generalization capability of deep object detectors on remote sensed trap images. Even with the application of the widely used geometric and texture-based augmentation techniques, the proposed methods can further increase the efficiency of object detector models. To demonstrate their efficiency, we used the Faster Region-based Convolutional Neural Network (R-CNN) and the You Look Only Once version 5 (YOLOv5) object detectors which have been trained on a small set of high-resolution, manually taken trap images while the test set consists of remote sensed images. The experimental results showed that the mean average precision (mAP) of the reference models significantly improved while in some cases their counting error was reduced to a third.