Research On Detection And Recognition Method Of Heterogeneous Small Object Based On Information Enhancement

The development of small object detection based on deep learning is an important research field to promote intelligent modern life.It is widely used in intelligent security,defect detection,intelligent agriculture and medical diagnosis,which greatly improves work efficiency.However,in complex scenes,object detection methods are faced with problems such as large changes in object scale,similar small objects and the background,and dense superposition of small objects,which makes it difficult to distinguish object detection methods from different scales,resulting in poor detection accuracy and generalization ability.To solve the above problems,this paper uses information enhancement,screening object resampling,multi-scale feature fusion and other methods to increase the training times of small objects and enhance the semantic information features,so as to enhance the information extraction ability of the model for small objects.The three main research contents of this paper are as follows:(1)Research on the information enhancement method of object location region based on pixel filtering.Due to the lack of visual information of small objects in the data set,less discriminative features can be extracted.This paper introduces a selective small object replication algorithm to resample the small objects in the data set.The number of small object samples is increased without increasing the number of images in the data set.(2)Research on Multi-scale Feature Fusion and Shallow Information Enhancement Method Based on Deep Learning.Aiming at the problem of incomplete feature information caused by insufficient shallow semantic information of small object,this paper proposes a multi-scale feature fusion and shallow information enhancement model based on deep learning.By combining Res Net-50 with the adaptation feature pyramid network,the object feature semantic information in the image is downsampled and extracted.Among them,the adaptive feature pyramid network is used to decouple the feature information of different sizes of objects,and assign more adaptive feature levels to small objects.Then,the extracted shallow semantic information and deep semantic information are laterally connected through deep learning technology,so as to realize the use of convolutional network and feature information splicing to extract the connection of semantic information of each layer of the object,and further strengthen the integrity of object information acquisition.The experimental comparison results demonstrate that the proposed algorithm model is capable of accurately detecting the smallscale image of distributed objects.(3)Research on detection and recognition method of small object based on information enhancement.Firstly,a pixel-filtered target location region information enhancement model is used to filter all the targets in the data set repeatedly,and the small target blocks are cut from the source image,copied and flipped and pasted to the original target image to directly obtain the combined amount of synthetic training data.Then,based on multi-scale feature fusion and shallow information enhancement model,the processed data set images are sampled and extracted to extract the object feature information,and the output feature map is processed by image size unification.After it is passed into the fully convolutional network,the heat map is obtained,in which the peak point position of each feature map predicts the width and height information of the object.Finally,the location size of the object box and the category of the object are obtained by regression of the central key point obtained by prediction.Compared with the detection effect of the original algorithm on the Safety Helmet Wearing Dataset(SHWD),the proposed algorithm achieves a mean average precision(m AP)of 89.06%and frames per second(FPS)of 28.96,an improvement of 18.08% m AP over the previous method.The detection results of small objects are also better than the original model and the current advanced detection model YOLOv4,which further proves the effectiveness of the proposed model for detecting small objects.