Font Size: a A A

Research On Target Detection And Recognition Based On OpenCV And Neural Network

Posted on:2020-01-05Degree:MasterType:Thesis
Country:ChinaCandidate:G L ZhaoFull Text:PDF
GTID:2428330572483637Subject:Control engineering
Abstract/Summary:PDF Full Text Request
The image processing technology has always been a hot research topic in the field of machine vision and artificial intelligence.It has important research significance in the fields of driverless,face recognition and artificial intelligence.The core idea of image processing technology is to use the computer-aided analysis method to detect and identify images and realize the semantic analysis of image sequences.Due to the uncertainty of noise in the image,the computational complexity of feature point extraction and matching,and the complexity of multi-classification problems,it brings difficulties to traditional detection and recognition algorithms.In this paper,the traditional target detection and recognition algorithms are studied in detail,and an improved image processing algorithm is proposed.Moreover,considering requirements of detection and recognition under different environmental backgrounds,the detection and recognition algorithms of neural networks are studied.An improved neural network algorithm is proposed,which realizes the precise positioning and accurate recognition of the target object,and overcomes the defects of the traditional target detection and recognition algorithm.Firstly,the main algorithms in the image preprocessing stage include mean filtering,median filtering,Gaussian filtering,etc.In this paper,the theoretical analysis and simulation of the above algorithms are compared,and the existing filtering methods can not meet the denoising effect under different environmental backgrounds.Therefore,an adaptive filtering algorithm is proposed to integrate the algorithm that dynamically changes the size of the filtering window into the denoising method,which effectively filters out the noise under different environmental backgrounds.In the target detection stage,the paper analyzes the static scenes,Dynamic target extraction method.From the inter-frame difference method,the background difference method and the optical flow method of dynamic target detection,an in-depth simulation analysis was carried out Considering that the multi-frame image sequence is composed of single-frame image sequences,the static target detection algorithm is studied.On the basis of this,the characteristics of the static target detection algorithm are found,and the existing feature extraction algorithm is improved.The effective elimination of redundant matching point pairs is realized,and the detection efficiency is improved;Then,in the target recognition stage,the traditional two-classification algorithm cannot meet the requirements of the actual classification due to the complexity of the background and the non-rigid variation of the target motion posture.Therefore,for the multi-objective classification problem,based on the in-depth study of the most commonly used SVM binary classification algorithm,the improved multi-level classification algorithm is proposed by means of the first verification and classification,which makes up for the shortcomings of the two classification algorithm.Maximum recognition of multi-target problems;Finally,this paper focuses on the advantages of a neural network algorithm and its improved algorithm.Considering the limitations of traditional algorithms and their improved algorithms,the advantages of neural network algorithms over traditional detection and recognition algorithms are compared and analyzed.At the same time,SSD and YOLO are verified.The characteristics of the two network models.The simulation results show that the existing network model has a poor effect on the detection and recognition of tiny objects.Therefore,combined with the existing model,the target confidence domain and back propagation algorithm are proposed to improve the neural network,which effectively improves the detection accuracy of the target detection and recognition and the robustness of the experimental results.
Keywords/Search Tags:Neural network, Multi-level classification, Feature extraction, Adaptive filtering, OpenCV
PDF Full Text Request
Related items