Online Specific Target Detection For Decimeter Level Optical Remote-sensing Images

Target detection in remote sensing images has always been a hot spot in the field of image processing. Decimeter level optical remote-sensing images could obtain texture and detail which meter level images lack. Target architectures are closer to characteristics of human ocular perception than SAR images, while the amount of data are much larger than remote-sensing images in the past. We must develop online specific target detection algorithm which is fit for the trait of decimeter level optical remote-sensing images to satisfy the needs of higher precision and speed.The source of this paper's thesis is based on "Processing System for Complex Targets in Optical Remote-sensing Images" which is cooperated with Chinese Academy of Sciences. With vehicle detection in decimeter level optical remote sensing images, for example, we focus on online specific target detection for decimeter level optical remote-sensing images, completed the following work:We proposed a rapid extracting target candidates'algorithm using combined confidence calibration method for remote-sensing images. We use combined confidence calibration method to calculate confidence of areas around candicate windows. Multiple high confidence sores are used to calibrate target location, avoiding the loss of target information effectively. We use binarized normed gradients to extract objectness feature, which has low computational amount and better generalization ability. Cascaded SVM framework could generate less amount of candicate windows than other alogrithms. Experiments on DLR Munich Vehicle dataset shows that the proposed method has the detection rate of 98.6% when generating 50000 candicate windows, reducing the amount of computation of the follow-up algorithms.Prime candidate windows have many negative samples, we proposed a deep convolutional neural network based on multi-threshold maximum gradient norms to eliminate negative data. Multi-threshold of the maximal gradient norm of three RGB channels of targets enhanced the target when shaded by trees or buildings, improved robustness of algorithm. The deep CNN contains 4 convolution layers and 4 pooling layers, extracting deep invariable features of samples, thus enhancing the learning outcomes. We use Caffe to training and testing neural network, improving efficiency and versatility. Our proposed deep convolutional neural network based on multi-threshold maximum gradient norms enhances the boundary of targets when objects are disturbed by complex noises, while is robust to the translation, scaling and rotation.We proposed a multi-scale spatial pyramid stochastic pooling based on multinomial distribution for deep convolutional neural network, achieving better detection results. Our neural network replaces max-pooling layer with stochastic pooling based on multinomial distribution, reducing the problem of overfit. Pyramid model with stochastic pooling extracts multiscale characteristic of targets, so the network could process different sizes of input images directly. With pre-trained model and fine-tuning, our network extract representations from the images in other datasets. Experiments on DLR Munich Vehicle dataset shows that our proposed algorithm has the precision rate of 93.3% and false alarm rate of 17.7% when recall rate is 95%.We proposed online specific target detection platform for remote-sensing images based on multi-core DSP TMS320C6678. We migrated the rapid extracting target candidates'algorithm and deep convolutional neural network to embedded platform of multi-core DSP TMS320C6678. By applying parallel optimization algorithm based on customization of CPU, we improved the performance of processing speed of detection method, which has the speedup of 6.01 and computing time of 0.11s, achieving online specific target detection for decimeter level optical remote-sensing images.We use the DR-#WIN evaluation metric, recall rate and calculation amount to evaluate target candidates'extracting algorithm. For the specific target detection algorithm, false alarm rate, precision rate and recall rate are the criterions used to evaluate performance. DLR Munich Vehicle dataset is collected from urban environment which is very difficult and challenging, is used to verify the detection performance. Comparing with the results of other algorithms such as HOG, LBP and MVC, our proposed algorithm achieved better detection rate, as well as less computation time, satisfying the requirement of online specific target detection of remote sensing system.