Low Light Face Detection Based On Convolutional Neural Network And Its Fpga Implementantation

With the development of artificial intelligence technology,face detection and recognition technology has received widespread attention,and related products of face detection are also developing in the direction of miniaturization and low power consumption.The use of image enhancement algorithm and face detection algorithm to achieve the preprocessing of low-light face images and the detection of faces in the image,the use of FPGA to achieve hardware acceleration of the algorithm,you can obtain a clear image under low light and assist the public security department to capture the face,which is of great significance to the face feature extraction of China’s security department.This thesis mainly studies the low-light face detection algorithm and its FPGA hardware implementation,based on the low-light face image data,using the homomorphic filtering algorithm to realize the adaptive parameter homomorphic filtering algorithm of the exponential function and the Gaussian high-pass filter function,the FAST image pyramid algorithm of image scale transformation and image splicing fusion using the MTCNN algorithm,and the hardware acceleration of the above algorithm through the FPGA,and the comparison of the results of the software and hardware experiments.The high speed and rationality of the hardware architecture design of the FPGA-based low-light face detection algorithm are verified.The main work of this article is as follows:1)An adaptive parameter homomorphic filtering algorithm based on weak-light face image is proposed.Aiming at the problem of multiparameter and empirical parameter value in traditional homomorphic filtering algorithm,the exponential homomorphic filtering transfer function with single parameter is used to replace the traditional multiparameter Gaussian transfer function.At the same time,the value of peak signal-to-noise ratio and structural similarity is quantified,and the optimal value of parameters is selected.The experimental results show that the homomorphic filtering algorithm based on parameter adaptive parameters has more obvious image enhancement effect than the traditional homomorphic filtering algorithm.2)A fast image pyramid algorithm based on MTCNN is proposed.Aiming at the problem of slow detection speed when the amount of input data in the MTCNN algorithm is large,the algorithm uses image scale transformation technology to obtain images of different sizes,and uses image stitching technology to stitch the transformed images to form a fixed-size face image,reducing the amount of input data,thereby improving the face detection speed.Experimental results show that the optimized MTCNN-based fast image pyramid construction algorithm is faster and more accurate than the MTCNN image detection algorithm before optimization.3)In order to meet people’s needs for miniaturization,low cost,low power consumption and other electronic products,a low-light face detection hardware implementation scheme based on FPGA is proposed.This scheme uses the abundant on-chip resources of FPGAs to quickly realize low-light face detection through design structures such as pipelines and off-chip memory.Experimental results show that the hardware processing effect of the algorithm is basically the same as that of the software,but the processing speed is faster and the resource consumption is lower,which achieves the real-time and low-consumption requirements,which verifies the rationality of the scheme.