Detecting drones at night with complex backgrounds poses a considerable challenge. To address this issue, we propose an enhanced variant of the YOLOv5 model, termed as the tiny feature and bidirectional feature pyramid network (BiFPN). This method incorporates efficient residual bottleneck (ERB) and efficient multireceptive pooling (EMRP) layers along with convolutional block attention module (CBAM). Utilizing ERB, the network achieves enhanced feature extraction through residual connections, whereas the EMRP layers incorporate multiple receptive fields, enabling the model to better understand and process varied spatial hierarchies within the data. Bi-FPN is added within the model's head to enhance feature representation and capture multiple features at various scales. To optimize model efficiency, cross-convolution replaces simple convolution, leading to a notable reduction in parameters. Furthermore, auto-anchor and auto-batch mechanisms are introduced to ensure efficient GPU utilization. Experimental evaluations conducted on a custom multiclass dataset illustrate a significant enhancement over baseline and state-of-the-art algorithms.
TF-BiFPN Improves YOLOv5: Enhancing Small-Scale Multiclass Drone Detection in Dark
IEEE Transactions on Aerospace and Electronic Systems ; 61 , 2 ; 5354-5361
2025-04-01
2056056 byte
Article (Journal)
Electronic Resource
English
DOAJ | 2023
|