Fractional Local Neighborhood Intensity Pattern for Image Retrieval using Genetic Algorithm
This work addresses the problem of improving image retrieval accuracy for users in computer vision applications, but it is incremental as it builds on an existing local pattern method.
The paper proposes a new texture descriptor called Fractional Local Neighborhood Intensity Pattern (FLNIP) for content-based image retrieval, which extends an existing method by encoding sign and magnitude information from fractional changes in local neighborhoods and uses a genetic algorithm to combine multi-resolution distances, showing significant improvement in precision and recall on four databases compared to recent state-of-the-art methods.
In this paper, a new texture descriptor named "Fractional Local Neighborhood Intensity Pattern" (FLNIP) has been proposed for content based image retrieval (CBIR). It is an extension of the Local Neighborhood Intensity Pattern (LNIP)[1]. FLNIP calculates the relative intensity difference between a particular pixel and the center pixel of a 3x3 window by considering the relationship with adjacent neighbors. In this work, the fractional change in the local neighborhood involving the adjacent neighbors has been calculated first with respect to one of the eight neighbors of the center pixel of a 3x3 window. Next, the fractional change has been calculated with respect to the center itself. The two values of fractional change are next compared to generate a binary bit pattern. Both sign and magnitude information are encoded in a single descriptor as it deals with the relative change in magnitude in the adjacent neighborhood i.e., the comparison of the fractional change. The descriptor is applied on four multi-resolution images -- one being the raw image and the other three being filtered gaussian images obtained by applying gaussian filters of different standard deviations on the raw image to signify the importance of exploring texture information at different resolutions in an image. The four sets of distances obtained between the query and the target image are then combined with a genetic algorithm based approach to improve the retrieval performance by minimizing the distance between similar class images. The performance of the method has been tested for image retrieval on four popular databases. The precision and recall values observed on these databases have been compared with recent state-of-art local patterns. The proposed method has shown a significant improvement over many other existing methods.