Aircraft Detection from Remote Sensing Images using YOLOV5 Architecture

Abstract

Aircraft is a means of transport and weapons that plays an important role in the civil and military sector for detection from remote sensing images. Although many algorithms have been proposed to improve this work but it is not very easy due to non-availability of structured datasets and annotations. However, the suggested algorithms have a number of flaws; for example, when applied to a remote sensing image, the algorithms would overlook several small-scale aircrafts. The problem has two primary causes. One is because the aircrafts in the remote sensing image are often modest in size, making detection difficult. The other issue is that because the backdrop of a remote sensing image is frequently complicated, the algorithms used to solve the problem are easily influenced by it. At present, the target detection methods supported by convolutional neural networks (CNNs)& YOLOV3 lack the sufficient extraction of remote sensing image information, lack of pre-processing steps and unfocused image augmentation which results in high missed detection rate and False alarms when facing small targets or object placed in the corners and boundaries. Aiming at the above questions, here is a proposed solution with the help of target detection model, using deep learning architecture that will be training on the custom-made model based on YOLOV4, YOLOV4 tiny, YOLOV5. In Proposed model YOLO series are chosen here because they have better detection techniques, more real-time applications present and the speed is relatively better than other series. The aim of the paper is to compare the model’s performance and choose the best working model on CGI plane Dataset. Basically, these models are detecting aircraft, creating annotation & bounding box to identify the target easily. These models have two stage Detectors which is used for achieving high accuracy in image recognition tasks. YOLOV4 have achieved the Mean Average Precision which is higher than YOLOV5 and YOLOV4 TINY on CGI Planes in Satellite Imagery in which firstly 400 images are present but after proper augmentation techniques there are 1000+ images present right now in which 800 images are in training set and rest 200+in test set. These models have less computational cost which reduces the execution time per time step during simulation as YOLOV5 takes 1.38 sec in 1 iteration which is way less than other working models.

Topics & Concepts