Geospatial Target Detection from High-Resolution Remote-Sensing Images Based on PIIFD Descriptor and Salient Regions

Geospatial target detection from visible remote-sensing images is considered as one of the most important issues in the analysis of aerial and satellite imagery. Development of remote-sensing techniques and enhancing resolution of images provide an opportunity to advance automatic analysis. The proposed methods of geospatial target detection have faced to a variety of challenges. Recently, local features are extensively used which play a very effective role in dealing with these issues. High intensity variations between targets and backgrounds in different images are the most critical challenges. Most proposed local feature descriptors are not able to deal with this amount of intensity variations and are accompanied with errors when facing them. In this paper, PIIFD descriptor has been applied to cope with intense intensity variations, as this descriptor is symmetrical against contrast. The proposed framework to automatically detect geospatial targets includes a supervised approach based on local features extraction and description and consists of three main steps including training, image searching, and geospatial target detection. In the training step, local features are extracted by UR-SIFT algorithm that properly matches with remote-sensing images and are described by the PIIFD descriptor. Due to the large dimensions of the extracted features, the SABOVW model has been used for quantization purpose. This model uses soft assignment of features to codebook, and the presentation provided by this model is used to train SVM classifier. In the second step, for the sake of computational efficiency, the salient regions of the image are detected by combining the saliency models, which reduces the image space to search the geospatial targets. In the third step, the salient regions are scanned by the sliding window approach and a descriptor of each position is generated. Eventually, the process of detection of geospatial targets will be carried out by applying the trained SVM model to each window. In order to evaluate the efficiency of the PIIFD descriptor, its performance is compared with descriptors such as SIFT, DAISY, LSS, and LBP. The results showed better performance of the PIIFD descriptor in detection of geospatial targets.