Target Detection In Remote Sensing Images Based On Strongly-supervised Part Models

With the rapid development of aerospace technology and sensor technology,the amount of available remote sensing data has constantly increased,while the quality of remote sensing image has been constantly improved.Compared to the traditional low-resolution remote sensing images,high-resolution remote sensing images usually have complex background,and more fine texture structures and spatial layout.Typical targets in high-resolution remote sensing images,such as aircrafts,ships and buildings etc.,generally have different colors and shapes,and their directions and positions often vary very much.Target detection in high-resolution remote sensing images can be widely used in military monitoring,resource exploration and other research areas.How to detect targets in high-resolution remote sensing images with huge volume of data has become an important research content in remote sensing science and technology.Based on this background,in order to describe the structure information of targets more precisely for detection,the strongly-supervised part model(SSPM)is proposed in this paper,and it is a multi-scale and multi-orient mixture part model.Before training,not only the bounding box of each object,but also all part bounding boxes belonging to each object and their semantics are all labeled in training sets.In the training stage,firstly,multi-scale feature pyramids of all training samples are constructed from the improved HOG analytic image features,and the model structure is initiated by minimum spanning tree structure based on these labeled information of object parts.Secondly,several part-based sub-models corresponding to different directions are trained by a latent Support Vector Machine(Latent-SVM)combining negative examples data-mining algorithm and stochastic gradient descent algorithm.Each sub model is composed of a coarse root filter,a set of twice resolution part filters,and a set of spatial location models.All sub-models are combined to form the proposed mixture model for target detection.Similarly,multi-scale feature pyramid of a test image is constructed at target detection stage.Then,the response scores is computed in feature pyramids by the trained mixture model using sliding window matching with dynamic programing and distance transform algorithm,and the detection results are obtained by setting thresholds to the response scores.Finally,a common non-maximum suppression algorithm is adopted to optimize the detection results.In this paper,a large number of high resolution remote sensing images are used to evaluate the performance of our proposed SSPM.The contributions of different parts of SSPM are analyzed.The performance trend of SSPM related to the number of parts?the ratio of labeled samples and the number of sub-models numbers are also analyzed.Moreover,in order to analyze the experiment results intuitively,the compared experiments are implemented by considering different directions and complexity.At last,the proposed SSPM is compared with the weakly-supervised part-based model?HOG-SVM and Exemplars-SVM methods.The experimental results show that the target detection accuracy of SSPM is 90.72%,and has been improved about 5% compared to the original part model which is better than both HOG-SVM and Exemplars-SVM,and the test results are quite satisfactory.Meanwhile,SSPM has also shown strong robustness to rotation and intra-class varieties,which can be seen from the experiments of test samples with different directions and appearances.