EfficientNet-YOLOv5: Improved YOLOv5 Based on EfficientNet Backbone for Object Detection on Marine Microalgae

Abstract

Object detection has been a popular task in deep learning. In marine microalgae detection, the dimension of the image in the marine microalgae is too large, but the object is too small compared with the images. Additionally, the number of images in each category differs greatly, which brings a great challenge to object detection. We propose EfficientNet-YOLOv5 to solve the two problems mentioned above. Based on YOLOv5, we improved the Backbone of YOLOv5 with EfficientNet. To further strengthen our proposed EfficientNet-YOLOv5, we offer a variety of useful tricks, such as offline and online data augmentation, multi-scale testing, multi-model ensembled, and LabelSmooling. Extensive experiments on marine microalgae have shown that EfficientNet-YOLOv5 has good performance. It also has very strong interpretability in the marine microalgae scenario. On the marine microalgae detection in microscopy dataset, we used only the EfficientNet-YOLOv5 model and obtained an online score of 44.73 percent. Compared with the baseline model (scored 42.38 percent), EfficientNet-YOLOv5 improved by 2.35 percent. In model ensembled, we received an online score of 50.683 percent using the ensembled model of EfficientNet-YOLOv5 and YOLOv5s for detection. Overall, our model obtained a considerable improvement in detection accuracy. Moreover, it also has excellent performance in inference speed and model size.