Research On Extraction Of Fast Local Image Features

Feature extraction is the prerequisite of many computer vision applications, such as image matching and registration, object recognition and tracking. Recently as the popularity of affordable mobile devices with cameras, computer vision algorithms have been improved and designed to run on these devices with lower memory capacity and computing power. At present, though some feature extraction approaches have obtained good results on image representation, they are suffering from high computation complexity which limits their applications in real-time operation. Developing extraction algorithms of local image features which are simple, fast and feasible has been the most urgent task to be done in the field of image representation.Based on the latest studies on fast local features, this dissertation proposed several new fast local features with regard to description methods of local image patches. Creative construction ideas were introduced in these methods, and experimental results on general datasets and practical applications have validated the performance of the proposed approaches. The main contributions of this dissertation can be concluded as follows:First, a construction method of local image descriptor directly utilizing interval division on order permutation is proposed to improve the robustness of the previous method. The simple descriptor has linear time and space complexity and it is inherently invariant to monotonic illumination changes. Experiments on Oxford dataset show that the proposed method is robust to minor image deformations. Its application on traffic sign recognition also exhibits the good performance and fast speed of the proposed descriptor compared to other existing features.Second, a local image feature invariant to rotation and monotonic illumination changes is presented based on the theoretical analysis of the relationship between order permutation and illumination invariance. This feature adopts a rotation-invariant fast corner detector, combined with local sampling patterns to achieve rotation invariance, and uses the description method based on partition on order permutation to achieve illumination invariance. The proposed feature shows a good robustness to rotation, illumination and image blurring in the tests of performance, and only consumes similar computing time as binary features. Experiments on video stabilization and panoramic image stitching further demonstrate its capability on practical image description.Third, in terms of the information compression ability of image quantization, a local feature based on image patch quantization is proposed. Two different quantization methods are introduced in descriptor construction, and a multi-scale keypoint detector is employed to adjust the orientation and scale of the proposed descriptor. Moreover, a robust estimation method of affine transform is presented and used in image registration experiments. Results show that the proposed algorithms are able to satisfy the demands of this application.Finally, specifically for scene classification applications, an improved census transform feature is proposed. Some ideas such as the center-symmetric local ternary pattern are introduced to carry out the census transform and build the new histogram feature. By these improvements, dimension of the histogram feature has been reduced and the speed of feature extraction and classification training has been accelerated remarkably. The experimental results on15-class scene dataset and8-class sport event database prove that the proposed method can speed up compared to the previous algorithm without any performance deterioration.