Research For Image Situation Awareness Based On Deep Learning

With the development of technology and explosive growth in information, building situation awareness system which can analyze data automatically and has a certain perception of environmental situation has become an important research topic. Situation awareness system has to identify the category and quantity of the targets in the image environment and analyze the state of the elements according to the multi-sensor information. After that, situation awareness system has to make some predictions of the trend of elements. In situation awareness system the discovery of the targets’ category and location is the key to achieve the overall function of the system. In a variety of sensor information image can visually show the targets in high accuracy and almost in real time, so image can be the main sources of information. But the discriminant analysis technology for image information has not yet reached the level of practical application. The complexity and uncertainty of the objects in image is the main reason contributing to technical bottleneck. In recent years the application of deep learning technology has made breakthrough progress in intelligent identification of image targets. Deep learning has led to revolutionary changes in computer vision field and has attracted attention of academics and industries. In this paper, we investigate the application of deep learning method in situation awareness system based on visible images and synthetic aperture radar images. We choose convolutional neural network(CNN) as the neural network model. Through the realization of convolutional neural network model and its extended model, we try to improve the identification accuracy and efficiency of image targets in situation awareness system in order to provide new solutions for elements perceived task in situation awareness system and reference for the advanced application of deep learning in this field.First of all, we complete classification of the objects in the scene by implementing convolutional neural network on GPU-based high-speed computing platform. We use the effective deep learning tool Caffe framework to implement CNN which refer to the outstanding models in ImageNet challenge depending on the configuration of our computing platform. We train the model by back-propagation algorithm and select Softmax classifier to identify targets’ categories in model’s output layer. We analyze effectiveness of the CNN model according to the classification result on Cifar-10 dataset. Finally, we verify feasibility of the CNN model used for situation awareness tasks through the experiments on visible images and synthetic aperture radar images.Secondly, we investigate the CNN model’s application in target detection based on the implementation of object classification. Through research and analysis we choose Fast R-CNN and Faster R-CNN two models to complete location-aware task. Using these two models complete image objects location awareness experiments both on visible images and synthetic aperture radar images. According to the results through the analysis of perceived precision and efficiency we choose Faster R-CNN model to complete targets location-aware task.Finally, according to the targets’ category and location information which given by the model we judge the main target and dense distributed target in the scene and give the text description of the scene. And according to system settings annotate the sensitive targets in image. Now we finish the first stage of situation awareness system.In this paper, we also investigate the possibility of situation awareness system in embedded platform. We configure Caffe framework and CNN model in Jetson TX1 developer kit. We proved that the CNN model can be realized in embedded computing platform through the model’s efficiency and results on mnist dataset.