| Summary: | 碩士 === 國立屏東大學 === 資訊工程學系碩士班 === 107 === In recent years, with the development of technology, image is inseparable from everyday life. Image recognition is a method of using artificial intelligence (AI) to analyze target images, and along with computers to automatically identify targets. There are many kinds of image recognition in daily life, for example: face recognition, fingerprint identification, pupil identification, aerial image recognition, etc. Because of the rise of UAV, aerial image recognition has become more and more popular, how to quickly and accurately identify large and abundant aerial image, is the most difficult topic for image recognition.
In this thesis, a RCNN (Region CNN) algorithm is used to identify the object in aerial image with a resolution of 4000*3000 pixels and without pre-processing. Firstly, the aerial image is segmented and reclassified, then a COCO format training set is created, and training the training set with the resnet101 model. Furthermore, it inputs the original aerial image, and marks the trained objects on the image. Finally, for local description of SIFT (Scale-invariant feature transform), and SURF (Speeded Up Robust Features), image matching of these two algorithms that are still widely used for image recognition, are compared with their recognition rate and recognition speed.
|
|---|
