Improving Small Pest Bird Detection in YOLOv5s for Autonomous Bird Deterrent Systems

Abstract

Granivorous birds are known to destroy grain crops in farms, and various studies are underway to find a solution to the problem. In recent studies, state-of-the-art deep learning technologies have been actively applied. However, image resolution has made detecting smaller pest birds a challenging task. Moreover, high-speed and low flight altitude bring in the motion blur on the densely packed birds, which leads to great challenge of object distinction. For that purpose, this paper presents an improved YOLOv5s model based on the YOLOv5 single-stage detector. The improved YOLOv5s model is proposed for application in bird deterrent systems where image background noise is high and identification of small birds is poor. To achieve this, the CSPDarknet backbone in YOLOV5s was replaced with DenseNet. Three convolution blocks and modules of the CSPbottleneck in YOLOV5s were also replaced with Transformer encoder blocks, and PANet in the original YOLOV5s neck was substituted with BiFPN. To further improve the performance of the improved YOLOv5s model, one additional prediction head was introduced for tiny object detection in the head. Both the original YOLOv5s and improved YOLOv5s models were trained using images from the Klim dataset. The dataset contains 1607 images for training, 340 images for validation, and another 357 images for testing. The test results on the Klim dataset showed an improvement of up to 4.8% in mean average precision when detecting smaller birds with the improved YOLOv5s at 50% Intersection Over Union, at the cost of just a 4 milliseconds increase in inference time. Based on a comparison with the original YOLOv5s model on the Klim dataset, the proposed YOLOv5s model outperformed the original model and achieved the highest performance in terms of accuracy (97.30%), area under receiver operating characteristic curve (93.78%), precision (98.54%), and F1-score (57.85%). The results showed that the modified YOLOv5s model is suitable for detecting small birds in various environments and consequently applicable in bird deterrent systems. Keywords— Deep Learning, Object Detection, Small Pest Birds, YOLOv5s