Research And Implementation Of Image Color Cast Correction Based On FPGA

Computer vision technology has been widely used in industrial production,urban construction,and smart security.Imaging equipment will be affected by ambient light,the spectral reflectance of the surface of the object,and the sensor’s own photosensitive performance.The image will have color changes due to different lighting conditions.difference.The purpose of calculating color constancy is to eliminate the influence of ambient light on imaging,so as to restore the truest color of the object and improve the reliability of the computer vision imaging system.At present,it mainly focuses on the research of the algorithm level,and less discussion about the realization of the hardware level.The ultimate goal of the algorithm is to be able to run safely and steadily in the hardware system,so it is very important to consider the realization of the color constancy algorithm on the hardware.This paper takes the color constancy under a single light source as the research object,and conducts the following research work from the algorithm layer and the hardware layer:First of all,several classic unsupervised color constancy algorithms are studied and implemented.Based on the color constancy theory,an evaluation system for the degree of image color shift is established from both subjective and objective evaluations,which is the basis for subsequent evaluation algorithms.Performance provides theoretical support.A horizontal comparison of several classic color constancy algorithms is carried out to provide a reference benchmark for algorithm performance analysis for subsequent research.Secondly,it conducts in-depth research on the color constancy algorithm based on deep learning,and uses convolutional neural network and transfer learning to establish an end-toend network model.Through the cluster analysis of the image data set,combined with the Softmax activation function,the neural network-based illumination estimation model is transformed into an illumination classifier.The experimental results show that while reducing the complexity of the illumination estimation problem,it also guarantees the effect of the color shift correction algorithm.In order to implement the illumination classifier,this article has implemented the selection,expansion and preprocessing of the training set,to the comparison of three classic network frameworks,and then to the model training,using visualization tools to try to explain how the neural network works in this problem.Finally,considering the realization of the color constancy algorithm on the mobile terminal,the lightweight Mobile Net was selected as the network framework,and the hardware acceleration of the network framework on the FPGA platform was completed,and the hardware transplantation of the algorithm was realized.When designing the CNN acceleration module,the design idea of multi-stage pipeline is adopted to maximize the parallel computing power of FPGA and optimize the transmission and storage of data.Design a complete imaging system,including image data acquisition,storage and display,and finally design an automatic white balance system equipped with artificial intelligence algorithms.