Font Size: a A A

Research On High Density Image Counting Based On Density Function Estimation

Posted on:2020-01-12Degree:MasterType:Thesis
Country:ChinaCandidate:G ZhangFull Text:PDF
GTID:2428330575463143Subject:Engineering
Abstract/Summary:PDF Full Text Request
With the rapid development of technology and fundamental theory research,intelligent image processing and analysis have become a very important and active research field.The studies of intelligent image analysis includes the classification,object detection and Semantic Segmentation in images and videos.Object counting is also an important research field in intelligent image analysis.Usually,object counting in low-density image can be realized by general object detection algorithm.However,the accuracy of the object counting in high-density image is still limited.Although in recent years,deep learning has been widely applied in high-density image counting,it requires GPU or a High performance computing platform.How to implement a high-density object counting algorithm with higher accuracy and robustness in normal platform or embedding system is also very useful.The main research contributions of this dissertation are summarized as follows:1.A density function model is constructed to estimate the real density function by using the 2D Gauss kernel function.Based on the high-density object counting dataset,a real density function is defined according to the every pixel of the training sample in the labeled data.By extracting the feature vectors of the image from each pixel,Gauss template convolution and the Gauss function are used to construct the model for real density estimation,through L2 regularization to control the parameter range,the optimal density parameter model can be obtained.Finally,the accurate high-density object counting can be calculated according to the integral value of density estimation.2.Based on the constructed density function model,this dissertation uses three different types of high-density object datasets for experiments.Two evaluation criteria are used to verify the accuracy and robustness of the proposed model in different high-density scenarios.The experimental results show that the proposed method has a higher counting accuracy comparing with the traditional regression counting method,.Compared with the method based on deep learning,the proposed model needs need less computation time and the proposed model is more suitable for computing platforms with lower performance such as embedded systems.
Keywords/Search Tags:Machine learning, target count, 2D Gaussian kernel function:density function parameter model
PDF Full Text Request
Related items