Image Classification Based On Deep Learning And The FPGA Design Of Feature Point Matching Algorithm

With the continuous development of artificial intelligence and computer vision,The Three-dimensional reconstruction technology based on multi-viewpoint images is also becoming a hot research direction,the Three-dimensional geometric model can be generated by processing the Two-dimensional images data obtained from ordinary cameras.Compared with professional Three-dimensional reconstruction instruments,Three-dimensional reconstruction technology based on multi-viewpoint images has the advantages of low cost,small limitation and less resource consumption.It has broad application prospects in Three-dimensional maps,Three-dimensional organs,Three-dimensional scenes of film and television,Three-dimensional reconstruction and protection of cultural relics and historical heritage.At present,there are still some shortcomings in Three-dimensional reconstruction technology based on multi-viewpoint images.This paper analyzes the shortcomings and puts forward some optimization strategies for the existing problems,The main work is as follows:1)In actual work and projects,in order to obtain the input data to meet the needs of Three-dimensional reconstruction,we usually use mobile phones,cameras or drones to collect multi-view images data of objects.However,the original images data collected by these manual methods can cause the problems of various types and unclear classification.Because of these reasons,the problem of mismatching between images data containing different information(non-ideal matching surface)will occur in feature points matching,and even the establishment of wrong correspondence between completely different images data will affect the accuracy of the final Three-dimensional model.To solve this problem,in this paper,the convolution neural network model(Convolutional Neural Networks,CNN)combined with the improved k-means clustering algorithm(K-means++)is proposed to classify and preprocess the original images data.By training and comparing the Alex Net,VGG16,and VGG19 models on the data set,the better VGG19 is selected as the images data feature extraction model.Clustering the extracted image feature vector representations to achieve classification preprocessing of images data and in the specific classification experiment,the accurate classification of images data is realized,which effectively improves the speed and accuracy of the subsequent feature point matching work.2)In the three-dimensional reconstruction technology based on multi-viewpoint images,the rate and accuracy of feature points matching are important prerequisites for the establishment of ideal three-dimensional model.The SIFT is usually used for feature points extraction and matching of images data.At present,the algorithm is mostly implemented by software based on CPU,however,The CPU-based implementation has some disadvantages such as slow speed,low efficiency,high power consumption and weak real-time performance in the actual Three-dimensional reconstruction work and projects.In order to solve the above problems,I proposed to use FPGA combined with HLS technology to optimize the design of images data feature points matching.I conducted some specific experiments on classified images data and compared the experimental results with the traditional CPU-based implementation.The realization rate is much higher than the traditional software method,and the expected purpose of high speed,low power consumption and real-time is realized.