My birds: utilizando inteligência artificial para a detecção de pássaros amazônicos

Abstract

Birds arouse human attention for their beauty and diversity, encouraging admirers to practice the recreational activity of spotting them and registering them in images. Sharing these records on citizen science platforms, such as WikiAves and eBirds, can significantly contribute to scientific research to understand and preserve these species. In this context, the sighting of birds can occur through feeders located in the gardens of residences, which contributes to the well-being of residents. The opportunity to record birds in feeders through webcams was noted. In addition, it was questioned whether it would be possible to automatically detect their species through deep learning, a technique widely adopted in computer vision. In this way, the present work aims to raise an approach based on deep learning to detect bird species that visit residential feeders. Therefore, the study had access to the feeder of a residence in the municipality of Santana, in Amapá, a region inserted in the Amazonian context. A publicly available dataset was collected, consisting of 940 images of birds and 1,836 notes distributed among 5 classes referring to the identified species. The images were obtained by extracting frames from the recordings made by the webcams, in addition to being annotated for the object detection task. The set was used to train different models of the Faster R-CNN type over two consecutive phases called preliminary and final. In the first, a smaller portion of the data was used to define the training configuration and a baseline. In the second, a single model called definitive was trained with all the data through the specified configuration, in addition to being compared with the baseline. The models were evaluated using the mAP, precision, and recall metrics, in addition to the use of confusion matrices. When considering the Intersection over Union at 50%, the definitive model achieved an mAP of 98.33%, a mean precision of 95.96%, and a mean recall of 98.82%. To the best of our knowledge, My Birds is the first work to propose detecting Amazonian bird species in residential feeders and raising an annotated set of these species. In this way, it is possible to visualize future works focusing on the automated collection of more images in other residences to contribute to data collection on these species.