Detection of a small target object in blurry images affected by affine distortions

Abstract

The paper proposes a novel and practically effective method for detecting, classifying, and estimating the coordinates of the image center of a small-size target object on a noisy scene, which is invariant to linear conformal transformations (rotation, shift, and scale). We consider a binary classifier that decides whether a particular part of the scene contains the desired image or only the background. The proposed approach implies an interactive procedure for finding an extremum of a function that approximates the likelihood function of the binary classifier. A two-step procedure based on the Nelder-Meade method is used to implement the extremum search. In order to ensure the robustness to noise and linear conformal transformations, both special training methods and the approach based on using an ensemble of classifiers, each of which corresponds to a certain scale, are applied in training the classifier. The author created a method for detecting a blurred image of a small-sized object in a scene that is distorted by correlated noise and proposes simultaneous estimation of the coordinates of the center of the target image. The method is robust to linear conformal distortions and has been successfully tested both on the artificial model and real images. The results of numerical study confirmed the robustness of the method to correlated noise of additive type and to linear conformal transformations. Within the framework of the proposed approach, the problem of constructing a confidence set for the coordinates of the target image center has been formally solved, and the efficiency of the obtained solution has been numerically investigated. The properties of the confidence set are formalized in the form of a theorem. The work also makes a comparison with the classical correlation-extreme method. If necessary, the proposed method can be easily generalized to the multiclass case. The method can be applied to machine vision systems, including online analysis of aerial survey data and to systems for video monitoring of the mechanical condition of complex technical equipment under conditions of strong meteorological disturbance.

Keywords

• image matching
• machine learning
• svm-classifier
• nelder-meade method