Research Of Feature Points Matching Method Based On BP Neural Network

As one of the main researches in computer vision, feature points matching is also one ofthe important elements of photogrammetry. So, exploring a new feature points matchingmethod combined with the specific application has very important significance.Based on the study of feature points matching principles and the existed matchingmethods, combined with the advantage of the neural network theory which can learn fromknown samples and generalize unknown samples, the paper proposes a new feature pointsmatching method based on BP neural network. The core idea of the method is: on the basis ofthe matching between two images, simulate the space mapping relationship of the featurepoint sets of them by determining BP neural network model, then complete the process offeature points matching under certain constraint criteria. However, the use of traditional BPneural network algorithm and single criteria lead to unsatisfactory matching results.Considering the lack of traditional BP neural network algorithm and single criteria, thepaper combines with the advantages of global optimization of genetic algorithm, furtherlyproposes an improved matching method using genetic algorithm to optimize BP network andadds the interaction constraint standard on the basis of the original constraint criteria. Thisapproach guarantees that the BP network gets global optimal space mapping relationship andeffectively filters mismatching points, which improves the accuracy and stability of thematching results.The above method is applied to photogrammetry system. Multiple experiments show thatthe method using only traditional BP neural network has lower correct matching rate andunstable matching results; while the improved algorithm can overcome these shortcomingsand improve the correct matching rate and stability, basically has achieved the intendedpurpose.