Rotation invariant estimation is an important and computationally difficult process in the real-time human computer interaction. Our new methodologies propose here for on-line image rotation angle estimation, correction and feature extractions based on line integrals. We reveal that a set of projection data of line integrals from single (fan-arc and fan-beam) or multi point sources (Radon transform) are employed for orientation estimation. After estimating orientation, image angle variations are altered to its principal direction. We further combine Boltzmann machine and k-mean clustering to obtain parameter optimized Gabor filters, which are used to extract non-redundant compact set of features for classification. The proposed method of fan-line, fan-arc and Radon transform are compared for real-time image orientation detection. Accuracy of classification is evaluated with the classifiers viz., back propagation, Hamming neural network, Euclidean-norm distance, and k-nearest neighbors. Experiment on a database of 535 images consisting of license plate and iris images. The viability of suggested algorithms has been tested with different classifiers. Thus, this paper proposes an efficient rotation invariant recognition for on-line images recognition.