Research On Image Pedestrian Detection And Rain Removal Algorithm Based On Deep Learnin

In the traditional monitoring information collection and intelligent vehicle-road coordination system,compared with ordinary targets such as traffic signs and vehicles,pedestrian detection is more challenging due to problems such as diverse pedestrian attitudes,variable scales,complex backgrounds and easy occlusion.In the rainy scene,due to the obstruction of the near rain line and the fogging effect of the distant rain,the image produces a large area of blur and a small range of deformation.Especially under heavy rainfall conditions,the presence of rain streak noise will seriously affect the image clarity,making pedestrian detection in a single image more difficult.For pedestrians,the aspect ratio varies greatly.Based on the Center Net object detection algorithm,pedestrian detection is defined as the prediction of the center key point.The improved deep aggregation network PDLA is used as the backbone network to extract features,which makes the downsampling process faster without affecting the details of the extracted features.Improve the thermodynamic layer that determines the key points,use the Gaussian ellipse as the Gaussian kernel,and derive the edge scale of the Gaussian ellipse.The improved Gaussian ellipse adaptively predicts the aspect ratio of the bounding box,which improves the detection ability of the algorithm for pedestrians with large differences in aspect ratio.The comparison of the detection results of a set of different characteristics with the benchmark label shows the adaptability of the algorithm in different pedestrian scenarios.The existing rain removal methods cannot effectively deal with rain patterns with different directions,uneven distribution and density in rain images.An image deraining algorithm based on information distillation attention recursive network is proposed,which performs feature extraction through efficient information distillation attention network,and uses multiple feature fusion attention module to further refine the extraction of rain pattern features.The overall network adopts a recursive strategy to gradually remove rain,prevent the loss of low-level information through the long and short-term memory module,and enhance the detailed recovery of images after rain removal.Experiments show that compared with other existing methods,the proposed method obtains higher indicators on three different simulation datasets,effectively removes rain lines while retaining a large number of image background details,and the overall visual effect is better.