Research On Blur-Invariant Feature Extraction And Blurred Image Recognition

The image will be blurred when there are relative motion, out of focus or atmosphere turbulence between camera and target. Image blur decreases the performance of the precision-guided weapons, which makes it important to develop a new method to improve target recognition performance under image blurred situation. Under such background, we propose a novel method for blurred image recognition by extracting blur-invariant features in the blurred image.In this paper, we firstly present the background and significance of the study of blurred image recognition. The current research status is also provided. Secondly, we research on both the texture and structure blur-invariant side. In texture blur-invariant feature extraction phase, we propose an improved LPQ feature, which is based on VLAD method. The feature uses dense sample method to extract the local blur-invariant information. VLAD is used afterwards to aggregate the sample features to the final improved representation. In structure blur-invariant feature extraction phase, we propose an improved HOG feature, which eliminates the weak gradient magnitude and is fused with the original HOG to improve its blur-invariance as well as keeping its discriminative power. At last, we fuse the texture and structure blur-invariant features together to form the final blur-invariant representation. The final blur-invariant feature is then used in the blurred image recognition.To test the generality and effectiveness of our proposed feature, we use different datasets for experiment, which include visible and infrared images. Different kinds of image blur are also used, which include the motion blur, out of focus blur, atmosphere turbulence blur and the actual aero-optical effects blur. Experiments show that our proposed feature can achieve better performance in both visual and infrared image recognition situation. It is also blur-invariant when images are blurred by the image blur shown above. The proposed feature can be used in the blurred image recognition. At last, we summarize the full paper and also analysis the future research directions.