The neural-SIFT feature descriptor for visual vocabulary object recognition

摘要: Recognizing the semantic content of an image is a challenging problem in computer vision. Many researchers attempt to apply local descriptors extract features from image, but choosing best type feature use still open problem. Some these systems are only trained once using fixed descriptor, like Scale Invariant Feature Transform (SIFT). In most cases algorithms show good performance, they do not learn their mistakes training completed. this paper continuous deep neural network feedback system proposed which consists adaptive bag visual words approach and classifier. Two initialization methods for descriptor were compared, one where it was on SIFT output randomly initialized. After initial training, propagates classification error classifier through entire pipeline, updating itself, also extract. Results that both increased accuracy substantially when regular able increase any further. The neural-SIFT performs better than itself even with limited number instances. Initializing existing beneficial lot samples available. However, there construct well-performing solely based feedback.