Image Pattern Recognition Using Phase-based Local Features and Their
Flexible Spatial Configuration
by Gustavo Carneiro
We propose a new image pattern recognition system that is applicable to several
computer vision tasks, such as long range motion matching and object
recognition. The main strength of our system is its ability to handle
substantial image deformations without significantly sacrificing the
expressiveness of the model representation. This system is divided into three
steps, namely: a) feature extraction, b) similarity search, and c) hypothesis
verification. The phase-based local feature proposed for step (a) is shown to be
distinctive and robust to 2-D rigid deformations and severe brightness changes.
The step (b) pairs similar model and test image features, producing the
correspondence set, which is usually densely populated with outliers. Hence, the
rejection of outliers from this set is necessary to reduce the number of
hypotheses to be verified in step (c). We propose two methods to reject outliers
that are robust to rigid and non-rigid deformations. Quantitative evaluations
for both the local feature extractor and the outlier rejection methods are also
provided. Comparison results produced by these evaluations show that our feature
is more robust and distinctive than state-of-the-art features proposed in the
literature, and our methods to reject outliers are more robust to 3-D rigid and
non-rigid deformations than the Hough transform, which is a common method used
to reject outliers. Finally, our last contribution is a probabilistic
verification for step (c) that uses local and semi-local similarities between
test and model images. The effectiveness of our system is tested in several
recognition problems.