Proposal of an Advanced Structure of YOLOX for Hornet Detection Accuracy Improvement

In this paper, an advanced backbone structure for YOLOX is proposed to obtain better detection accuracy in small object detection such as hornet by replacing CSPLayer with ShuffleLayer. By this replacement, numbers of convolution operation are reduced in each layer of the backbone. This can conserve spatial information of small objects in each layer and through layers in backbone, reducing processing time. In order to evaluate the proposed method, four types of experiments were executed such as mAP comparison for our hornet dataset, another mAP comparison for the standard dataset VEDAI dedicated small objects, generalization test for RTMDet, and detection speed between the default YOLOX model and the proposed YOLOX model. As a result, the first mAP under 50% IoU condition for the hornet dataset showed 86.21% and 87.35% for the default and the proposed, respectively. The experiment, mAP test for the standard VEDAI, represented 47% and 41.7% for each model and also showed better accuracy by 5.3%. In the generalization test with RTMDet, the proposed model showed similar or higher accuracy according to IoU. In addition, in terms of speed the proposed ShuffleLayerbased backbone was faster than the default by 1.35 times due to reduced convolution parameters. Thus, experiments above verified that the proposed backbone structure for YOLOX can be effectively utilized to enhance accuracy and inference speed in real-time detection for small objects.