Abstract:
Correlation Patter Recognition Filters have been explored by several researchers to improve performance for recognition of objects. An innovative supplement provided by logarithmic translation as a pre-processing block, reinforces the capabilities of these filter to handle variations in viewing angles and size during object recognition process. With advent of logarithmic correlation filters, computation complexity is considerably reduced for the training process, because we do not have to train correlation filters separately for achieving invariance to different distortions. Rather we get this functionality in a single training instance without creating multiple banks of filters.
Logarithmic versions of correlation filters show a distinguishing performance in object recognition especially when an input image carries multiple distortions. However, some very efficient recognition algorithms have been discovered in the recent past, therefore further improvement in terms of detection accuracy and clutter defiance of correlation filters is considered essential at this stage.
Many feature-based methods have evolved and the same can be utilized to strengthen performance of correlation-based recognition. Some of the well-known and time-tested methods like active contours, shape context and Scale Invariant Feature Transform can provide a real benefit, if combined with the correlation-based recognition. This research encompasses combining the strength of classification methods depending upon features and Correlation Filters, to achieve a failsafe object recognition application that can be used in several fields.
Incorporation of features within design of correlation filters is carried out while training of the filters. Subsequently the same information is utilized during testing and classification process to obtain better results duly authenticated by correlation pattern recognition and selected feature-based technique.
This novel combination of two different domains helps to suppress side lobes generation in correlation output. The side lobes are caused by clutters and the same can distort the main correlation output and dominate main correlation peak thus making classification decision ambiguous.
