학습 데이터셋에 따른 딥러닝 모델의 말벌 검출 정확도 비교 (Comparison of Vespa Detection Accuracy for Deep Learning Models According to Training Dataset)

서지정보

· 발행기관 : 한국양봉학회
· 수록지 정보 : Journal of Apiculture / 38권 / 3호 / 223 ~ 232페이지
· 저자명 : 곽희진, 권영재, 이철희

초록

Training data configuration is critical for object detection based on supervised deep learning.
Namely, the characteristics of training data should be very similar to an actual test environment to raise expected inference accuracy. However, a Vespa object size in pixels in a natural capture environment is not constant and is smaller than the Vespa object size of images in a training set. This study compares the inference accuracy of deep learning models YOLOv5, YOLOX, and YOLOv7 according to training datasets. As training data, the three types of datasets were prepared as follows. First, A basic dataset, composed of five species of Vespa and one bee, is produced for the Vespa training data set. Next, The 0.3% dataset, in which Hornet object size is approximately 0.3% ratio to the whole image size in pixels, is prepared using the basic dataset for tiny object detection. Finally, images are selected from the basic and the 0.3% datasets in the same proportions in a Mixed dataset. After configuring three types of training datasets, the three deep learning models above were trained using the three training datasets, the basic, the 0.3%, and the Mixed dataset, and calculated the training and test mAP. In cases where the training and test data environments are similar, YOLOv7 demonstrated the highest mAP at 95.4%. However, in a test result experiment for actual environments using trained weights by the basic dataset, the mAP@50 scores are 30%, 14%, and 85% for YOLOv5, YOLOv7, and YOLOX, respectively. That is, YOLOX, an Anchor-free model, is overwhelmingly excellent. The organization of the training dataset is essential to match the inputs of the actual detection environment to obtain the best accuracy in object detection, and YOLOv7 is the best model for a tailored training dataset among state of the art models.

영어초록

Training data configuration is critical for object detection based on supervised deep learning.
Namely, the characteristics of training data should be very similar to an actual test environment to raise expected inference accuracy. However, a Vespa object size in pixels in a natural capture environment is not constant and is smaller than the Vespa object size of images in a training set. This study compares the inference accuracy of deep learning models YOLOv5, YOLOX, and YOLOv7 according to training datasets. As training data, the three types of datasets were prepared as follows. First, A basic dataset, composed of five species of Vespa and one bee, is produced for the Vespa training data set. Next, The 0.3% dataset, in which Hornet object size is approximately 0.3% ratio to the whole image size in pixels, is prepared using the basic dataset for tiny object detection. Finally, images are selected from the basic and the 0.3% datasets in the same proportions in a Mixed dataset. After configuring three types of training datasets, the three deep learning models above were trained using the three training datasets, the basic, the 0.3%, and the Mixed dataset, and calculated the training and test mAP. In cases where the training and test data environments are similar, YOLOv7 demonstrated the highest mAP at 95.4%. However, in a test result experiment for actual environments using trained weights by the basic dataset, the mAP@50 scores are 30%, 14%, and 85% for YOLOv5, YOLOv7, and YOLOX, respectively. That is, YOLOX, an Anchor-free model, is overwhelmingly excellent. The organization of the training dataset is essential to match the inputs of the actual detection environment to obtain the best accuracy in object detection, and YOLOv7 is the best model for a tailored training dataset among state of the art models.

참고자료

· 없음