Research Of Adherent Noises Detection Algorithm Based On Edge Defocused Model

Human beings are entering the information age, With the computer comes into our lives more and more deeply, we have mastered some solutions to the problem of the computer vision tasks through the application of applied mathematics knowledge, such as indoor surveillance systems, driving assistant system. The advanced driving assistance system (ADAS) based on computer vision technology analyzes image information to tell driver the information of traffic conditions, and gives some warning of the potential danger in time. However, due to the on-board camera is in the outdoor environment for a long time and vulnerable to the adherent noise on the lens shade (such as mud, leaves, etc.), which causes the system unable to get complete and true information of outside world, this will cause the error of judgment. Therefore, adherent noises detection play a crucial role for the whole system to help it work normally. This paper focuses on the adherent noises detection problem of on-board camera.Because the adherent noises detection is done in the process of driving, and the shape, size and position of the adherent noises are all not sure, this brings certain difficulties to the adherent noises detection. This paper proposes a adherent noises detection algorithm based on edge defocused model through application of the edge and region features of adherent noises to detect it.This paper is based on the defocus imaging model and the edge of the adherent noises that is blurry due to defocus blur phenomenon,extracting the degree of defocusing at the edge position to describe the edge feature of the adherent noises. In addition, this paper considers the correlation of gray inside of adherent noises in space, extracting the characteristic parameters of gray level co-occurrence matrix (including contrast, correlation and homogeneity) to describe the regional characteristics of the adherent noises.In this paper, according to the edge features of adherent noises, we select the candidate edge points and use these points as the seed points, so the candidate regions of adherent noises are generated by the iterative region growing algorithm. Then, with the help of the fusion feature parameters to combine the information of three characteristic parameters, the region features of the adherent noises are applied and the probability model of fusion feature parameters to do a further validation for the candidate regions of adherent noises is built. So we can get the result based on single frame. Finally, in order to reduce the error detection and missing detection in single frame detection, this article uses multi-frame images to verify the candidate region under the condition of multi-frame point by point, so we can get a more accurate detection results then the detection based on single frame.Through experiments under different conditions and scenes, we can find the adherent noises detection algorithm which proposed in this paper is suitable for vehicular environment, free from the interference of complex background, and has high precision, good effect. Our algorithm is able to complete the real-time adherent noises detection in the process of driving or parking.