Research On Object Recognition Technology Based On Visual Attention And Local Invariant Feature

The object recognition is the fundamental purpose of the visual system. How to recognize the object from complex scenes are more important and difficult problems. The local invariant features are invariant to scale and rotation, and has more strong robustness for affine distortion such as viewpoint changes, light changes and noise. Hence it can complete the reliable recognition between the same objects which is obtained with different angles. However, the dimensions of the each local invariant feature descriptor generally choose high dimension, and as the image size increase, the number of descriptors increases linearly, thus increase the amount of calculation, it is difficult to meet the real-time requirements of object recognition system.The object recognition system requires as little as possible amount of calculation, because of it have the limited computing resources and the high real-time demand. Visual attention is a very effective tool which is able to solve the problem of the computing resource consumption. Human visual attention can quickly select a small amount of information related to visual task under complex scenes, so leading the human visual attention to target recognition system based on local invariant features has broad application prospects.In recent years, the related studies have proposed many object recognition methods, combining bottom-up visual attention with the local invariant feature. However, the existing methods are still not satisfactory. The main reason is that the quality of saliency map itself is not high enough; the typical local invariant feature method is not suitable for the matching of low resolution ROIs; application modes and methods of visual attention mechanism also are not mature enough.This paper fully summarized and compared the existing saliency detection algorithms and local invariant feature methods. On this basis, fast but robust saliency detection and local invariant feature methods are proposed, and object recognition method based on combination of visual attention with local invariant feature is designed.In this paper, the main work and innovation points are as follows;(1) Along the ideas of frequency-domain information and frequency tuning, this paper proposes two kinds of quick and accurate detection algorithm of object region; IHFT and IMSSS. Aimed at the long running time of HFT algorithm, IHFT transforms the convolution operation into point-to-point multiplication operation to accelerate the processing on frequency-domain information. Aimed at the saliency map fusion operation of MSSS algorithm, IMSSS performs the selection processing of feature map based on the entropy calculation of each feature map instead of add operation of three color feature map. The theoretical basis of these two algorithms are comparatively single, the detection of object region under complex background has some limitations. In order to complement each other’s advantages of IMSSS and IHFT, this paper presents a fusion algorithm based on IHFT and IMSSS, which can adapt to more complex scenes and improve the adaptability and detection accuracy of algorithm.(2) Along the idea of BRISK method, this paper proposes a novel logarithmic-Spiral local invariant features the method, referred to as LOS-K. This paper proposes a novel local invariant feature method to improve the matching performance of image patches with the low-resolution and small size. The location, scale and orientation of local invariant features are directly estimated from an original image patch using a Log-Spiral sampling pattern for local invariant feature detection without consideration of image pyramid. A Log-Spiral sampling pattern for local invariant feature description and two bit-generated functions are designed for generating a binary descriptor. Extensive experiments show that the proposed method is more effective and robust than existing binary-based methods for image patch matching.(3) Aiming at the problem of target recognition under complex scene, this paper proposes a target recognition method based on the Attention-Recognition Fusion Model (referred to as ARF model). This method recognize ROIs obtained by the combination of the bottom-up saliency map and top-down Intimacy map using local invariant feature method. By introducing a novel top-down attention component, our method selects the ROI and ignores the meaningless clutters, thence reduces the complexity of object recognition based on local invariant features. For the object recognition problem of synthetic test images and complex scenes, our method can greatly improve the recognition speed without reducing the recognition rate.