hierarchical artificial neuronets
A. V. Timofeev, O. A. Derin
The approaches used for image recognition, are divided into algorithmic, analyzing the scene by algorithms and based on the description of the objects on the prototype, and on "neural networks", using artificial neural network, trained on sample images and do not require the algorithm. Each of these approaches has its advantages, in particular, an algorithmic approach allows the use of accumulated human expertise and neural network – does not require a mathematical model of recognizable complex scene and the algorithm creation. The combination of these approaches and their advantages of traditional methods is impossible.
Contemporary trends of the theory of image processing is a joint analysis of problems of image recognition and generation. A method for constructing a hierarchical neural network (HNN) recognition of objects in multi – images (MI), whose structure inverted structural diagram of a multimedia system – a computer graphics processor (GP). Described the structural scheme of the HNN, performs a reverse sequence of actions as compared with GPs, described algorithms of its elements. This structure was implemented in the language C++ program as filter of the DirectX multimedia subsystem of the operating system Windows. HNN has been tested to detect MI CCTV systems on one railway station at Saint-Petersburg in order to detect suspicious persons in the waiting room. The system was developed under the R&D performed SPIIRAS (St. Petersburg). HNN showed high robustness with respect to the background image and the probability of detecting objects – namely, the probability of recognition of people / vehicles / animals is not less than 0.8, including control the trajectory of motion – not less than 0.75, the probability of detection of objects left behind – not less than 0,6.
Based on the proposed structure of the HNN is possible to construct three-dimensional video sensor of the surrounding space. The combination of the detected objects in neighboring frames in the MI allows for moving the robot to estimate distances to these objects and build a relief of the surrounding space. In this case, the stereobase is the distance traveled by the robot, i.e. stereobase is dynamic and can be changed by correcting the speed. It is proposed to divide the segmentation of the HNN on the supporting performing segmentation of the next image from scratch, without relying on the results of previous calculations (based on clustering by K-mean), and intermediate segmentation results of the previous correction algorithms (based on an algorithm of competitive learning neural network). An example of the results of the prototype soft-hardware device three-dimensional view, which provides automatic detection of surrounding objects and circumvent obstacles the mobile platform. The conclusion about the appropriateness of the proposed structure of the HNN for the selection of moving objects in the MI.