Optimization of the keypoint density-based region proposal for R-CNN

In areas such as computer vision, the content recognition of an image is a topic of interest in applications such as search engines, biometric security and autonomous cars, among others, since the computer must recognize all the objects that an image can have, which arises as the challenge of localizing and classifying different objects inside a single image in an efficient way. In recent years, this challenge has been approached with the use of region-based convolutional neuronal networks (R-CNN) which are systems that learn to recognize different objects by their representation in a series of images. The proposal of regions is essential for the performance of R-CNN when locating the individual objects of the image with accuracy and in the shortest time. In this article we propose a modification to a method for region proposal based on the density of SIFT like feature points that describe the objects within the image. The selection of regions is made through a decision based on the values of the cumulative distribution function of the normal distribution constructed using points density. The obtained results show a significant reduction in the processing time required for the localization of objects; having slight variations in the classification accuracy with respect to using methods such as KDRP and selective search.