由于半导体工艺的限制,图像传感器列并行读出电路的系统结构很容易出现列与列之间的失配,从而导致输出图像在垂直方向亮度不均匀,表现为明暗变化的竖条纹,即FPN噪声,其严重影响了图像质量。本设计基于矩匹配法对其进行改进,提出了一种针对图像传感器的FPN噪声去除算法。首先在系统上电时,让图像传感器采集随机不同的300帧图像,然后算出所有图像的列像素均值和累计像素均值,当前列均值减累计像素均值即为该图像传感器当前列的校正值。本设计以Zynq系列ARM+FPGA异构SoC作为实现平台,OV6946作为图像传感器,使用Vivado HLS进行输入/输出接口都为AXI-Stream的视频流处理算法IP的开发,并固化到Zynq的FPGA中。经过实验,本算法在Zynq7020平台、50 MHz的系统时钟下,处理OV6946摄像头输出的400×400@30 f/s视频流延时约为3.3 ms,保证了视频流输出的实时性,较好地实现了FPN噪声的去除效果,并且算法的应用具有普适性。

Due to the limitations of semiconductor technology,the system structure of image sensor column parallel readout circuits is prone to column to column mismatch due to process deviations,resulting in uneven brightness of the output image in the vertical direction,manifested as vertical stripes with varying brightness,which is also known as FPN noise.FPN noise seriously affects image quality.This design is based on the moment matching method and improves it,proposing an FPN noise removal algorithm for image sensors.Firstly,when the system is powered on,let the image sensor collect 300 randomly different frames of images,and then calculate the column pixel mean and cumulative pixel mean of all images.The current column mean minus the cumulative pixel mean is the correction value of the current column of the image sensor.This design uses Zynq series ARM+FPGA heterogeneous SoC as the implementation platform,OV6946 as the image sensor,and Vivado HLS for the development of video stream processing algorithm IP with AXI-Stream input and output interfaces,which is solidified into Zynq's FPGA.After experiments,this algorithm processes 400×400@30 fps outputs from the OV6946 camera on the Zynq7020 platform with a 50 MHz system clock.The video streaming time is about 3.3 ms,ensuring the real-time output of the video stream.The removal effect of FPN noise has been well achieved,and the application of the algorithm is universal.