Study On Feature Extraction Technique Based On The Perceptive Invariability

Feature extraction is the one of the important techniques in pattern recognition. How to extract the steady and reliable feature is the key step to accomplish the recognition. Images will be affected by the displacement, scale, rotation, complicated varying illumination. This thesis mostly considers the illumination invariant in the feature extraction technique, and gives attention to the displacement, scale, rotation invariability. Our aim is to extract the feature with displacement, scale, rotation, illumination invariability. The thesis theoretically studies the feature extraction technique based on the perceptive invariants. Apply and expand Hu moment invariants by the experiment, then gain the displacement, scale, rotation invariability. This thesis sums up the methods on how to eliminate the effects from the complicated varying illumination. A model-based Local Normalization Technique and Relative Gradient are adopted for illumination invariance. The LN algorithm works by illumination model to estimate the effects, and simulates illumination effects with the multiplicative noise and the additive noise. The Relative Gradient algorithm improves the gradient, extracts the inner character as the illumination invariant. Adopt the YaleB faces to do the experiment; the two algorithms both obtain the better results. But the LN algorithm is better than Relative Gradient. The thesis puts forward an algorithm based on Local Normalization Technique and improved Hu moments. With the LN algorithm and improved Hu moment, we can extract the perceptive invariability, the recognition result is better than directly extracting Hu moments. Experiments also show that our method is effective and fast.