Research And Implementation Of Image Haze Removal On Zynq Heterogeneous Platform

Atmospheric environment deteriorated in recent years.Droplets and dust suspended in the haze air interfere with the scatter and absorption of light,resulting in problems of optical sensors such as color distortion,contrast reduction and dynamic range of the image decrease.The haze images are difficult for information recognition which restrict traffic navigation,military reconnaissance,meteorological observation and target recognition.Statistics show that there is low brightness in some local areas of the haze free image.The dark channel image haze removal algorithm can recover the blurred image information by calculating the atmospheric light and transmittance in the atmospheric scattering model,which has a very good effect.However,this process takes a long time and can only reach the speed of 1.1fps on the embedded Cortex A9 processor,which is difficult to apply to practical fields.The heterogeneous computing technology based on CPU,GPU,DSP and FPGA is more and more used for special requirements of data volume,operation time and power consumption.It has become a hot direction in the field of embedded system.The embedded programmable chip of Zynq architecture not only has advantages in implementing complex control logic and scheduling task resources,but also has the characteristics of hardware acceleration and dynamic reconfiguration.The combination of the control processor and the programmable logic device can effectively improve the system integration,reduce the power consumption,and speed up the upgrade of the hardware equipment.In order to improve the efficiency of the dark channel image haze removal algorithm,I analyze the calculation and I/O amount of the algorithm.And a new system is designed to allocate the computing task according to the performance characteristics of the general purpose processor and the programmable logic device.Through positioning the bottleneck of the algorithm,the measures of parallel,pipeline and ping pong are used.I encapsulate a hardware acceleration core,which improves the calculation speed of atmospheric light and transmittance in the algorithm.Then a high throughput data path with low delay is built between the processor and the acceleration core through the AXI interface.And the video buffer processing architecture is constructed.Finally,the Linux operating system is set up on Zynq.The driver of the algorithm acceleration core is used for the interaction between the software application and the hardware.And the image acquisition,processing and display system based on OpenCV and DRM is built.Based on the basic principle of the dark channel image haze removal algorithm,a complete set of image processing system is constructed.The hardware method is 5.6 times faster compared with the software method.The speed is increased to 140%.At the same time,the volume and power consumption of the device are reduced,which has high practical value.Through the efficient data path between the heterogeneous cores,the developers can quickly adopt other processing algorithms,making the system not only a common platform for ac-celerating image processing,but also widely used in automatic driving,speech recognition,medical diagnosis and other fields.