Small Object Detection And Application In Complex Sense Based On Deep Learning Algorithms

With the rapid development of computer vision technology,object detection has been widely used in many fields such as public security,smart cities,and traffic control.At the same time,the development of deep learning algorithms makes the performance of object detection systems far superior to traditional algorithms,and further promotes the application of object detection.However,since there are a large number of small pixel targets in most application scenarios,such targets have a small amount of information,and the features are not obvious.The detection of small pixel targets is still a challenge in the field of computer vision.This article first proposes two small object detection algorithms in common scenarios;second,collects small target data sets,and applies the algorithm to face detection data sets,public drone scene data sets,and experimental collection data sets,calculate the objective indicators of small object detection;Finally,the algorithm is embedded in the intelligent terminal hardware system,and the corresponding software is designed and written to realize the application of small object detection algorithm.The main research contents are as follows:(1)An improved small-face detection algorithm SG-FACE based on Generative Adversarial Networks(GAN)is proposed.The algorithm first uses fuzzy targets in the image due to small pixels to generate high-resolution targets by generating an adversarial network,and then uses the new targets as input to the face detection network,through "concave" feature fusion structure and multi-scale detection structure.The results show that,compared with the traditional face detection,the SG-FACE algorithm small face detector has high accuracy and can detect a large number of targets missed by the traditional face recognition algorithm.The detection algorithm has high reliability and can be effectively used in security or other systems in comlex scenes.(2)A multi-scale fusion small pixel object detection algorithm YOLO-D4 based on YOLOv3(You Only Look Once version3)network structure is proposed.YOLO-D4 uses deep learning networks and clustering methods to fuse feature information extracted from 4 different scales,and realize the detection of different layers of semantic information.YOLO-D4 uses K-means clustering on the training set to find the best prior.At the same time YOLO-D4 adds a small target detector that uses four times downsampling features to fuse high-level feature information,it can fully integrate the background information of small object targets and semantic information,These changes improves the model's ability to detect small object features.Applying this algorithm to the mid-altitude and high-altitude small target scene data set.The results show that,compared with YOLOv3,the YOLO-D4 algorithm proposed in this paper improves the mean average precision(mAP)by 2.16%and the mid-and high-altitude view angle mAP by 2.76%.The mAP increased by 5.51%in the night vision scene and increased by 2.66%in the foggy scene.(3)Realization of small object detection system based on intelligent terminal.First,transplant the YOLO-D4 algorithm to the deep learning intelligent platform terminal;second,write the supporting software;at the same time,collect experimental data sets,perform data amplification,and improve the performance of the actual application scenario;finally,perform real-time on the intelligent terminal object detection.The results show that the small object detection system based on the intelligent terminal has a detection frame rate of 15.8FPS,which can meet real-time requirements.YOLO-D4 can effectively detect small pixel targets in different environmental scenarios,and has good generalization ability and practical application ability.In summary,for small pixel targets in complex scenes,this paper proposes a small face detection algorithm SG-FACE based on an improved generation confrontation network and a multi-scale fusion small pixel object detection algorithm YOLO-D4 based on the YOLOv3 network structure.The algorithm is transplanted to the intelligent terminal,the supporting software is written.At the same time we collect the actual scene data.Software and hardware system are realized.The theoretical and practical application development of the machine vision related algorithm is completed.