Automatic object detection in aerial image using bent identity-convolutional neural network and fine tuning algorithm

• Published in: Multimedia tools and applications Vol. 81; no. 7; pp. 9713 - 9740
• Main Authors: Lalitha V.P., Rangaswamy, Shanta
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.03.2022; Springer Nature B.V
• Subjects: Algorithms; Artificial neural networks; Background noise; Computer Communication Networks; Computer Science; Data Structures and Information Theory; Deep learning; Feature extraction; Fields (mathematics); Image analysis; Image classification; Image segmentation; Localization; Machine learning; Multimedia Information Systems; Neural networks; Object recognition; Remote sensing; Satellite imagery; Spacecraft recovery; Special Purpose and Application-Based Systems; Urbanization; Fine tuning algorithm; Aerial image; Bent identity; Convolutional neural network; Remote sensing
• ISSN: 1380-7501; 1573-7721
• DOI: 10.1007/s11042-022-11948-7

In RS (Remote Sensing) imaging, one of the most important tasks is Object Detection (OD) which aims to extract the object location and class information. Object detection is termed the highly advanced approach, which is helpful in RS image analysis, scene and image content understanding. In recent years, the researchers have made extensive efforts to present an effective OD model with the limitation of detecting specific classes such as roads and buildings in the RS images. For capturing aerial images, satellite sensors are utilized, and these images are affected by various factors like background noise, variation of viewpoint and interferences etc. To provide accurate detection of objects, an effective method is needed in analysing aerial images. However, this work presents an automatic OD using BI-CNN_FTA (Bent Identity-Convolutional Neural Network Fine-Tuning Algorithm) to detect more object classes relating to urbanization applications. The overall framework of proposed method is divided into five phases-Pre-processing, Object Localization, Segmentation, Feature Extraction and Classification. The input image is pre-processed by Advanced Dual Domain (ADD) filtering. Next, the objects are localized with the help of Grid Guided Localization (GGL) which draw an imaginary box around each object in an image. The segmentation process is done with the help of an improved tree Markov (ITM) random field model. Further, various features like texture, geometric, color descriptors and Zernike moments are extracted. Finally, the object classification is performed by BI-CNN_FTA. During this stage, each object class are categorized based on urbanization applications and learning loss is optimized. The proposed method is executed in the PHYTON platform. It has obtained improved performance in terms of Mean Average Precision (mAP = 85.5%), F1-score (83.78%), Detection rate (97%) and Precision-Recall curve when compared with other deep learning networks on the DOTA dataset.