Abstract:
Drone growth in recent years has brought up a number of opportunities as well as
challenges in a number of sectors, including aviation safety, security, and animal
protection. Accurately identifying and classifying drones from birds is a significant
difficulty for efficient airspace surveillance and collision avoidance systems. This paper
describes a deep learning-based method that uses the SqueezeNet architecture to
accurately categorize birds and drones. The labeled images in the dataset, which was
divided into training, validation, and test sets, were taken from
Roboflow.Comprehensive data preprocessing methods, such as augmentation and
normalization, were used to improve model performance. SqueezeNet and a customized
Convolutional Neural Network (CNN) were the two models created and assessed [1].
The SqueezeNet model exceeded the bespoke CNN model, which had an accuracy of
97.51%, with an amazing accuracy of 99.51%. SqueezeNet is a lightweight and efficient
model whose outstanding performance highlights its applicability to real-time drone
identification applications.To make sure the models were reliable and resilient, a lot of
evaluation metrics and visualizations were used[2]. The study comes to the conclusion
that the SqueezeNet-based classification method provides an effective way to
differentiate drones from birds, boosting surveillance capabilities, safeguarding wildlife,
and boosting aviation safety.This study also emphasizes how crucial careful model
evaluation and data pretreatment are to getting the best results from deep learning. This
technique has ramifications for many different applications, such as improving security
standards, aiding in the conservation of species, and reducing aviation hazards.
Prospective avenues for investigation encompass broadening the dataset, refining model
architectures, executing practical deployments, and tackling moral and legal implications
linked to drone detection technology. This work advances drone detection systems and
highlights the ability of deep learning models to tackle challenging classification
problems in a variety of settings.
