Research And System Implementation Of Video Small Object Detection And Object Attribute Recongnition Based On Deep Learning

As a discipline that allows computers to understand and process images like humans,computer vision has its own unique charm.With the continuous advancement of human science,it plays an increasingly important role in human life.For decades,the method of video object detection has improved the detection effect continuously.More and more detection methods have greatly improved the video object detection,these algorithms solve the video object detection problem in a certain scene to some extent,but more of them can only be limited to a specific scene,lacking robustness to algorithms in some scenarios,and training a good model requires a lot of tuning techniques,not universal.Among them,the small object detection of video is more difficult for video object detection,so it is a very challenging and difficult subject to develop an algorithm that has a good effect on video small object detection.At the same time,in order to analyze the small target information more accurately,this paper introduces the attribute recognition of the object,but for the video object,we usually regard people as the focus of attention,so this paper puts the focus of attribute recognition on the identification of pedestrian attributes and obtains certain result.Based on the deep learning technology,this paper adopts convolutional neural network to complete the analysis and research of video small object detection and pedestrian attribute recognition.The main work is as follows:(1)This paper studies the more important object detection models in recent years,mainly studies three types of models,one is the object detection of traditional methods,such as DPM,and the other is based on candidate region extraction models,such as R-CNN series.There is also a model based on regression,such as YOLO and SSD.By analyzing the effect of these models on the object detection in the image,the rationality of improvement based on YOLO is explained(2)Self-built small object detection dataset.In the case of the missing small object dataset,this paper collects and supplements the aerial dataset of the drone,and builds its own dataset for video small object detection.(3)Re-clustering the self-built small object dataset to ensure reasonable distribution and accelerated the convergence of the self-built dataset,and adopting multi-scale fusion,similar residual structure and adding dilated convolution to improve the recall rate of the small object detection.(4)Research and implementation of pedestrian attribute recognition.This paper introduces the multi-attribute joint mechanism into the pedestrian attribute recognition task,and designs the pedestrian multi-attribute identification network,and constructs the PMA network to complete the identification of pedestrian gender,age,clothing,and other attributes.In this paper,we use the correlation between attributes,the deep network to extract complete features,and the weight of uneven distribution of attributes to identify pedestrian attributes.