Adaptive Component Discrimination Network for Airplane Detection in Remote Sensing Images

Abstract

Airplane detection and recognition in high-resolution Remote Sensing Images (RSIs) remains a challenging task due to the factors of multiple view angles, multiple scales, multiple orientation etc. This paper proposes an Adaptive Component Discrimination Network (ACDN) for airplane detection and recognition in RSIs, which focus various scales from global to local, making full use of the overall contour as well as the dominant component features of airplanes. Firstly, a Standardization Processing Module (SPM) is proposed for image projection conversion and resolution uniform to alleviate the confusion of different types of airplane in different resolution images. Secondly, the Rotatable Bounding box-based Pyramid Network (RBPN) is utilized to extract candidate airplane coordinates and categories. Further, an adaptive aircraft component discrimination method is established for confusing few-shot airplane targets recognition, which consists of a Target Orientation Adaptive Adjustment Module (OAAM) and a Component Discrimination Module (CDM). OAAM obtains airplanes with same orientations by predicting the orientation of the slices and rotating them adaptively. All the uniformed slices are then fed into the CDM for dominant components detection, which corrects the target preclassification results, improving the classification performance. Experiments conducted on the 2020 Gaofen Challenge demonstrate the efficacy and superiority of the proposed method.

An Approach for Aircraft Detection using VGG19 and OCSVM

“…Feature extraction stands as a critical phase in airplane detection, as the model's efficacy significantly hinges on the quality and pertinence of the extracted features [35], [36]. This stage is dedicated to capturing pertinent visual characteristics from images that aid in distinguishing airplanes from other objects or backgrounds.…”

HTD: Heterogeneous Task Decoupling for Two-Stage Object Detection

“…Based on the development of general object detection and oriented object detection, FGOD in remote sensing has received increasing attention more recently [7], [8], [9], [10], [11], [12], [13], [14], [15]. Comparing with previous finegrained recognition work based on classification task [61], [62], [63], [64], FGOD requires simultaneous localization and fine-grained recognition.…”

“…cessful applications in remote sensing [4], [5], [6], it no longer meets the new demand for fine-grained recognition. In recent times, FGOD in aerial images has garnered widespread attention from the research community [7], [8], [9], [10], [11], [12], [13], [14], [15]. FGOD is a multi-task learning problem comprising foreground and background (FG/BG) classification, box regression and fine-grained recognition.…”