Lightweight Deep Learning Model-Based UAVs Visual Detection

Unmanned aerial vehicles (UAV) are often known as drones and are designed to introduce considerable benefits for privacy security and surveillance systems applications. However, the common use of drones demonstrates great threats to public security and personal privacy. Drone objects usually have mixed vision with bird objects due to their behavioural and physical similarity. Lightweight deep learning is a paradigm of machine learning that stands to efficacy and compactness for improving the performance of deep learning models. It is beneficial and suitable to apply with low-powered, Internet of Things (IoT), and edge devices, which have constrained resources. In this paper, a lightweight deep-learning model for drone vs. bird visual detection in colour images is introduced and investigated. The proposed model encompasses customized layers of convolution neural network CNN for feature extraction purposes as well as dense layers for binary classification tasks (drone/birds). The training process was accomplished using forward and back-propagation workflows to find the drone object in the input images. For training data, a pre-processing stage is performed to prepare the drone’s color image dataset for the input layer with various image resolutions, shapes, backgrounds, and localization in the input images. In the validation phase, an optimal set of hyperparameters was selected and investigated accurately to enhance the classifier performance. Layer pruning and network quantization techniques are applied in the initial network design phase and network training phase respectively, meaningfully reducing its parameters and memory usage for realizing a lightweight deep network. The experimental results and tests demonstrated that the increase in network layers or the number of the used filters, with each convolution layer in the proposed CNN structure, will not enhance or support the detector performance with higher accuracy. By increasing the trained drone/bird images, we have motivated the proposed model to be more converged and get a lower learning/validation error rate in addition to higher classification accuracy in the testing phase. Based on various experiments with public image datasets (drones/birds), the performance evaluation metrics have been exposed to a detection accuracy of 97.5