图像特征的一个重要分支就是纹理特征,它体现了不同图像和物体的形态、大小、分布、方向等重要参数,对图像特征识别起到决定性因素。但是纹理特征提取的过程十分复杂且计算量巨大,为了解决这个难题,提出了一种在现场可编程逻辑门阵列(FPGA)平台下实现纹理特征提取新方法。首先对基本图像特征算法做了并行化的优化,从算法的数值范围和表示精度两个角度,做了相应的分析和误差控制,从而适应FPGA的运行。然后对FPGA的数据流传输提出了一个高效率的解决办法,该方法对其中的主要模块采用了流水线优化,并采用寄存器配置模式,从而在线地修改参数,适应不同的图像大小和卷积核等环境变量。结果表明,在同等功耗条件下,可以达到10倍于CPU的性能,达到了快速提取特征的目的。
The image feature is an important branch of texture feature, which reflects the different images and object shape, size, distribution, direction, and other important parameters and plays a decisive factor on image characteristics recognition.But the texture feature extraction process is very complex and time cost.In order to solve the problem, a new method to extract texture feature based on FPGA is implemented.First, the texture feature extraction method is optimized with parallel algorithm, then the error is analyzed and controlled based on numerical range and representation accuracy, so the method can operate on FPGA efficiently.Also, a method to improve the data stream transmission on FPGA is designed, which employs pipeline optimization on main modules and register allocation model.The system on FPGA can modify parameters on-line to adapt for different environmental variables, such as image size, convolution kernel and so on.The results show that the proposed model extracts image texture feature up to ten times faster than CPU under the same power consumption, and it is an ideal system to fast extract image texture feature based on FPGA.