Research On Target Detection Algorithm Of Improved Faster R-CNN Haze Image

Target detection is an important basic research in the field of computer vision,and it is also an important foundation of image content understanding.With the indepth study of machine vision technology,Through a variety of intelligent algorithms to accurately identify and locate the target in the image.However,under the condition of haze,the resolution of the image taken outdoors is not high and the contrast is low,which results in the difficulty of target detection and low recognition rate.So it is very important to study the enhancement of haze image and target detection.According to the formation mechanism of haze image,the haze image enhancement and target recognition are carried out.The specific research work is as follows:(1)A multi-scale convolution network is proposed to realize end-to-end image defogging by cross communication.Firstly,based on the principle of atmospheric physical model,the fog image restoration is carried out,and the fog image data sets of different degrees needed for deep learning network training are made.Using convolutional layer cross communication to reduce network parameters can also enhance the nonlinear ability of network structure.The experimental results show that the defog network can effectively remove the fog on the basis of preserving the image information.(2)A method of haze image enhancement based on spatial segmentation is proposed.A 3 × 3 receptive field is designed by the concept of Gauss function.The mapping relationship between spatial and frequency domain is obtained.The image is divided into low frequency,medium low frequency,medium high frequency and high frequency regions,which are enhanced by gamma calibration,msrcp,MSR and top hat methods respectively.Finally,they are combined.Compared with the traditional image enhancement algorithm,it avoids the filter to filter the noise and preserve the high frequency information of the original edge part of the image,that is,the image itself,to enhance it.(3)The parallel positioning optimization structure is proposed to improve the traditional Faster R-CNN.The input of the network is divided into four parts,and the initial frame part is added to synthesize five-dimensional position information.The parallel five-way CNN structure is used to extract the target position information and complete the mapping calculation of individual candidate boxes,so as to achieve the positioning optimization.The experimental results show that it can effectively identify and frame the target object,reduce the situation of missed detection and false detection,make the object location more accurate,and the recognition rate increases by 8.6% on average.This subject has a certain innovative significance and practical value,and plays a positive role in the realization of target detection in haze image.