| Summary: | As the largest species of birds in Antarctica, penguins are called “biological indicators”. Changes in the environment will cause population fluctuations. Therefore, developing a penguin census regularly will not only help carry out conservation activities but also provides a basis for studying climate change. Traditionally, scholars often use indirect methods, e.g., identifying penguin guano and establishing regression relationships to estimate the size of penguin colonies. In this paper, we explore the feasibility of automatic object detection algorithms based on aerial images, which locate each penguin directly. We build a dataset consisting of images taken at 400 m altitude over the island populated by Adelie penguins, which are cropped with a resolution of 640 × 640. To address the challenges of detecting minuscule penguins (often 10 pixels extent) amidst complex backgrounds in our dataset, we propose a new object detection network, named YoloPd (Yolo for penguin detection). Specifically, a multiple frequency features fusion module and a Bottleneck aggregation layer are proposed to strengthen feature representations for smaller penguins. Furthermore, the Transformer aggregation layer and efficient attention module are designed to capture global features with the aim of filtering out background interference. With respect to the latency/accuracy trade-off, YoloPd surpasses the classical detector Faster R-CNN by 8.5% in mean precision (mAP). It also beats the latest detector Yolov7 by 2.3% in F1 score with fewer parameters. Under YoloPd, the average counting accuracy reaches 94.6%, which is quite promising. The results demonstrate the potential of automatic detectors and provide a new direction for penguin counting.
|
|---|