Research On Multi-feature Fusion Of Aerial Photography Image Recognition Taken By UAV

Unmanned aerial vehicle(uav) technology has been developed rapidly as a new aerial platform in recent years and widely used in the field of low-altitude photogrammetry. At present, hotpot and difficulty in image processing collected by uav mainly focus on preprocessing and splice, etc. Research on automatic recognition of uav aerial image combining digital image processing and pattern recognition method is of great significance as well as enrich the application range of uav and strengthen the effective use of aerial images.This paper puts forward the viewpoint that make image recognition by using multi-feature fusion approach in consideration that the accuracy isn’t high using single image feature to do with the image recognition. Because there is lot of image feature vector in the method of multi-feature fusion it is difficult to determine the feature weight. The idea of using classification method of neural network and adjusting the threshold through the forward feedback constantly to obtain the ideal classification results are carried out.The main work of this paper is as follows: introduces the related technology such as image recognition, image preprocessing, image segmentation, image feature extraction, feature fusion and research status of the uav aerial image processing both in domestic and abroad.For there is always lens distortion and aerial images are easily affected by weather factors, image preprocessing method is under deep research and the method of lens distortion correction, image defogging is realized after analyzing image lens distortion correction and image defogging based on dark transcendental.Because there is a wide range of aerial photography and numerous features categories, the aerial images are divided into many blocks according to certain pixel size based on analyzing image Segmentation method in threshold, region and edge detection.In this paper, the color, texture and shape feature extraction of image are also studied aiming at the problem multi-feature fusion. This paper analyzes the imagecolor moments, color histogram, color correlogram and pointing at the characteristics that extracting method dimension is too high, the HSV color space is nonuniform quantization, finally get the aerial image of 22 dimensional color feature vector.Based on texture features analysis the space correlation method, GLCM texture features, Tamura texture features, Gabor wavelet transform, texture feature extraction method, the aerial image based on GLCM texture features eight dimensional texture feature vector extraction; Based on form feature extract and analysis the shape description method, the Hu invariant moments and realize aerial images of seven Hu invariant rectangular shape feature extraction.The method designs the method flow using multi-feature fusion of aerial image.By classifying the image color histogram, GLCM texture, Hu invariant moment single feature through the BP neural network classification we get single color feature of the aerial image recognition. This method performs good recognition rate, but the identification effect of single texture and shape feature is very poor.Image recognition result is obtained by neural network classifier after fusing two or three of features like image colors, texture and shape features. Results show that after multi-feature fusion improves the aerial image overall recognition rate; by contrast, we found that the shape and texture features for single feature recognition are very low and the object classes is key features which can greatly affect the recognition precision of this kind of objects.