An Improved Randomized Local Binary Features for Keypoints Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Randomized local binary features are crucial for keypoint matching in methods like RandomForests, RandomFerns, BRIEF, ORB, and AKAZE.
• Existing methods exhibit limited randomness, primarily in sampling positions, potentially hindering feature quality.

Purpose of the Study:

• To enhance the quality of binary feature spaces by increasing randomness in the basic sampling operator.
• To introduce and evaluate a novel Randomized Intensity Difference (RID) operator for improved image patch observation.

Main Methods:

• Developed the Randomized Intensity Difference (RID) operator with randomness across grids, position, aperture, weights, and channels.
• Integrated various RID operators into RandomFerns and RandomForests classifiers.
• Applied the method to both grayscale and color images for recognition tasks on image and video datasets.

Main Results:

• Demonstrated that a completely randomized sampling approach (RID operators) generates a higher quality binary feature space compared to traditional RIT features.
• Experimental results confirmed the excellent performance of the RID feature method in recognition tasks.
• Proposed new evaluation criteria for robustness and distinctiveness to analyze randomization effects on binary feature spaces.

Conclusions:

• The proposed RID operator significantly improves the quality of randomized local binary features.
• Increased randomness in the sampling operator is key to generating superior binary feature spaces.
• The RID operator offers a versatile and effective approach for keypoint matching and recognition in computer vision.