Research And Application Of Small Object Detection Based On Deep Learning

In recent years,advances in deep learning have made computers perform better than humans on many visual tasks,which has made computer vision become a research focus in the field of computer science.As a basic computer vision task,object detection is one of the most important tasks.Small objects are the objects which have fewer pixels in the image.The information reflected by small objects is so dificient that detection methods designed for regular objects can't detect small objects well.Also,it's difficult to label the training data.Based on the analysis of object detection methods,this paper studies the small object detection based on deep learning and applies the improved algorithm to practical applications.In convolutional neural networks,low-level features contain more detailed infor-mation and are more capable of expressing small objects than high-level features.In view of this fact,this paper proposes a multi-scale detection algorithm based on Faster-RCN-N.The algorithm adds new region proposal networks to the original Faster-RCNN structure,so that the model can simultaneously detect small objects with features of d-ifferent scales.Meanwhile,to solve the problem that high-level features are not good at representing small objects,this paper proposes a feature enhancement algorithm based on CGAN.The algorithm uses CGAN to generate the residuals between features of s-mall objects and regular objects,and uses the generated residuals to strengthen features of small objects.In this paper,experiments are carried out on two datasets for small object detection and one dataset for regular object detection.The experimental results show that the multi-scale detection algorithm based on Faster-RCNN and the feature enhancement algorithm based on CGAN can improve the accuracy of small object de-tection while keeping the accuracy of regular object detection.To solve the problem that training samples of small objects are difficult to collect,this paper crawls images containing large objects and weak supervised information from the search engine,and uses these images to train model after they are transformed by down sampling and up sampling.In addition to studying small object detection based on deep learning,this paper also applies multi-scale detection and feature enhancement to the vehicle flow analysis system which is based on UAV video.In the system,this paper uses small object de-tection algorithms to do vehicle detection and calculates the deviation of the calibrated reference object to eliminate the jitter.On the basis of vehicle detection,Hungary algo-rithm is used to establish the association between vehicles in different frames.Finally,the vehicle flow is counted by the calibrated counting line.The method proposed in this paper can be used for vehicle detection in UAV video and the system can be used to analyze UAV videos captured from various shooting angles.