Parallel Processing of Infrared Images Processing in Thermo Vision Systems

摘要

Thermo vision are used in military, police custom traffic control, industrial and other specific applications for collecting and processing thermo visual information from infrared images. The problems arise in the steps of implementation of the developed methods and algorithms in real time practical applications of thermo vision systems. In surveillance and security thermo visual systems one of the most practical goals is the moving objects detection and tracking in infrared images captured from a thermo vision camera. The input infrared images are usually separated and processed in small blocks with an appropriate and chosen shape (for example rectangular) and size (for example 8x8). In conventional hardware or software implementation of infrared image processing algorithms the blocks are processed consecutively or in series and the achieving the real time processing is not always possible. The advances in powerful parallel computer graphics and image processing for computer vision and computer games applications with the developed graphical processing unit (GPU) and Compute Unified Device Architecture (CUDA) offers for GPU-based computing a powerful development framework integrated with high level parallel programming languages like C or C++ languages. Graphical processing units (GPU) are devices designed to exploit parallel shared memory-based floating-point computation. They provide memory access speeds superior to those of commodity CPU-based systems. These features to update in parallel the model variables every iteration compared to other solutions like programmable logic, integrated circuits, custom shared memory solutions, and cluster message passing computing systems make GPUs attractive in real time image processing and especially in this article for infrared image processing applications. Here is proposed to exploit the ability of parallel processing and the high-speed memory access of graphical processing units (GPU), which is essential in the real time applications with neural networks in most of the infrared image processing applications. In most applications of infrared image processing with neural networks the processed algorithms work sequentially by a CPU, which means only one neuron is updated at a given time. As a result the performance degrades quickly with the increase in network size and connectivity. This is especially the case for large connectivity, since sequential processors need to iterative over every connection for each neuron. To speed up the operation, supercomputers or distributed computers are normally used for large-scale neural network simulation. But these solutions incur high cost. Traditional CPU architectures are not designed for parallel processing. To avoid this problem in real time infrared image processing applications a suitable type of neural network is proposed to use the spiking neural network (SNN) implemented in graphical processing unit (GPU) and Compute Unified Device Architecture (CUDA). The example is presented for real time infrared image processing applications like moving objects detection and tracking in infrared images in surveillance and security thermo visual systems.