Research On Image Recognition Algorithms For Mobile Devices

As the Smartphone become increasing popular, the performance of mobile devices has been improved greatly, even in some aspects, can emulate or surpass the low end PCs. Therefore, it become possible to implement the complicated computer vision algorithms on mobile devices. With the continued development of technology application on mobile devices, such as visual reality, visual navigation and so on, the demand of computer vision algorithms application on embedded mobile devices become increasing urgent. Image recognition as the basic research topic of mode recognition and computer vision fields. The intense research of this subject, especially the image recognition algorithms of mobile devices has a significant research value.Because of the high recognition accuracy and good stability, image recognition which based on image matching has received more and more attention from researchers. While the image matching is inseparable from the image feature extraction and matching search, how to achieve faster image feature extraction as well as precise and quick match have become a key part of application for image recognition on mobile devices. Although researchers have been came up with various image feature extraction algorithms, it is still hard to fulfill the requirements of accuracy and real-time performance. Therefore, this issue still should be researched on in the future.This thesis aimed to achieve real-time image recognition on mobile devices, so I studies in three aspects:image local features extraction, feature description and fast search match among features; and finally makes a full implementation on Android platform and systematic reviews the performance of algorithm. The main research and achievements of this thesis are as follows:(1) Due to that AGAST feature point detection algorithm has poor robustness and does not have the scale and rotation invariance, this thesis combined the scale space theory and rotational invariance theory to improve the AGAST algorithm. The improved feature point not only included the scale and direction information, but also achieved the higher position accuracy, robustness has been improved significantly.(2) This thesis proposed that using binary descriptor based on study to match images can well overcome the contradiction of uniqueness and real-time performance for descriptors in feature description period. It also made the corresponding adaptive improvement for AGAST descriptor in order to suit the improved algorithm for detection of Binboost, and presented the improved scale and rotation invariant Binboost descriptor.(3) For low efficiency feature matching process in linear search and the existence of the "curse of dimensionality" in nearest neighbor search, the approximate nearest neighbor query algorithm was raised to achieve the high speed match for feature points, and carried out by Hamming space LSH algorithm.(4) The real-time image recognition framework has been designed and implemented in Android platform.