부산대학교 도서관

Aerial image classification is of great significance in computer vision, having applicability in geological study, agricultural, urban planning, military, etc. It is a challenging task due to the high variability in image appearance. Existing methods often fail to accurately classify aerial images in challenging conditions, such as when there is significant variation in illumination, viewpoint, occlusion, object size and scale, or image quality. To address these challenges, this work presents a novel approach for aerial image classification that leverages layer-wise feature extraction and concatenation. CNNs are useful for image classification because they can automatically learn to extract relevant features from images, which are then used to train a classifier to predict the image class. CNNs have proven effective for image classification, but their last layer features are high-level, so the report suggests extracting multi-layer features. VGG-19 is a convolutional neural network that is 19 layers deep. The proposed approach extracts features from multiple layers of a deep convolutional neural network (CNN) and concatenates them to produce a comprehensive feature representation of the input image. This comprehensive feature representation is then used by a fully connected layer to classify the image. The proposed approach is evaluated on a Aerial Image Dataset (AID) dataset of aerial images that includes a wide range of challenging conditions. The results are indicative that the presented approach achieves state-of-the-art performance on the dataset, outperforming existing methods by a significant margin. The work presented in this paper has the potential to be used for a variety of applications, such as land cover classification, object detection, and disaster monitoring.