Construction Of Feature Invariants And Its Application In Object Recognition

Extracting or construction of the invariant features is one of the key technologies in the field of pattern recognition and computer vision. It will be much easier for classification or recognition to extract the original shape features efficiently and farthest without the affect of various geographic transformations (such as rotation, translation, affine transform, etc) and without the influence of image degradation introduced by the bad condition or environment in the real scene. The moment invariant set is an important feature set from which various powerful invariants can be constructed according to some special requirements. More and more attentions are paid to this technology. However, the invariant feature set available for recognition is relative small. The functions of each invariant need to be concluded and their application scope need to be extended. More theoretic analysis and meticulous experiments in the field are also need. Under the condition of this research background, in-depth investigations are carried out in invariant feature extraction and its applications in this paper. The original research productions in this paper are listed as follows. (1) The general rules for constructing the rotational moment invariants are concluded, and a convenient method---trigonometric function for generating rotational moment invariants is proposed with many new general invariant formulas generated, and a big rotational moment invariant set is suggested. Based on the new method and the notion of sub-polynomial we have found, the multi-filter moment invariant generator algorithm is presented. Through this algorithm, large numbers of repeated sub-polynomial computation are avoided, which can accelerate the generation and the computation of moment invariants. The formula for the computation of moments of a region from its contour is presented too. (2) The geopmetric meaning of affine transform parameters and the relationship between rotational and affine transform are discovered. Some complemental researches are carried out on the construction of affine moment invariant. Several new affine moment invariants are introduced and applied successfully in recognition of surface images in 3-D space even when the depth information is unknown. The notion of affine moment vector is defined, by which the registration and normalization of the affine deformed images can be implemented successfully. (3) The general formula of n-D normalized central moment is defined, which is invariant to translation and scale of n-D object. Two new 3-D moment invariants Crot3,1 Crot3,2 are introduced which is also invariant to the rotation of 3-D object. These invariants can be applied to recognize 3D graphics. (4) A study is carried out for the transform rules of each moment invariant under the condition of image being blurred and the integral region is big enough. For the first time, through the method of complex moment, a blurred moment invariant subset is provided. Experiments have shown that this subset can be useful tool for recognizing the rotational, translated, scaled and blurred object images. (5) The outer contour and inner hole in an image scene can all be extracted by traditional contour tracing algorithm, and can not be easily distinguished. Aim at this circumstance, a discriminant to distinguish outer contour from inner hole is provided. (6) A general method to normalize the magnitude of moment invariant value of different order and degree is suggested, which can counterpoise automatically the influence of different features on classifier. Some key technologies in BP neural network for image recognition are suggested and employed, including estimation of training progress, confidence estimation, designment of the neural network, setting the step and initial value for training, etc. Experiments are performed for recognizing the scaled rotational or affine deformed objects in the image, or by its contour. All conclusions listed above are supported by experiments and theoretical analysis.The researches will provide important theoretical basis for the designment and implementation of an automatical target recognition system.