Neural Network Based Blur Measurement And Segmentation For Partially Blurred Image

The captured image can be partially blurred very often, possibly due to the weather, camera sensor, illumination and the relatively fast motion between the camera and the object. Effective extraction of the partially blurred image of the blurred the image can be used for image restoration or image fusion, etc. In this thesis, we mainly study the blur measurement and segmentation of the partially blurred image. The main contents are as follows:(1) Local blur measurement based on BP neural networkThere are two limitations among existing blur metrics:1) It is easy to detect of flat textured clear areas as blurred areas; 2) most existing blur metrics are only applicable to the global or only suitable for measuring motion or defocus blur. We propose a BP neural network based image partial blur measurement method. In this method, a new unified blur feature based on all singular values and non-zero DCT coefficients is presented, which measures the sharpness from both spatial and frequency-domain. Different singular values reflect the distribution of different scale information and vary differently after blurring. The number of non-zero DCT coefficients depicts the information loss in the high frequency domain. Their combination can effectively capture the blurring effect in the flattened textured area. Then, BP neural network and support vector machine are presented as the classifiers to effectively measure the blur. Qualitative and Quantitative experiments with single or multiple partially blurred images show that the method can effectively distinguish between the blurred and texture-flattened clear areas.(2) Image semantic based blurred region segmentationThe existing blur measurement methods are difficult to accurately segment the blurred region. Thus, this thesis proposes a semantic segmentation method to extract the blurred region, with the observation that the partially blurred areas are generally semantic-consistent as patches. This method consists of three steps:1) a bilateral filter is applied first to filter out the initial blur measurements by the method discussed in the previous chapter. The bilateral filter can effectively remove the noise and preserve the edge of the blur region, which accordingly help improving the segmentation precision; 2) a deep neural network based semantic segmentation method is used to segment the partially blur image, and can segment the specific semantic object; 3) if the measure of blurred region mainly concentrated in a semantic object, the object is considered as blurred regions, or else the blurred region is obtained through the threshold segmentation. The experimental results show the effectiveness of the proposed method.