Fast rotation and scale invariant object detection based on ensemble classifier : application to deformable cloth pattern recognition

다국어 초록 (Multilingual Abstract)

The core of many Computer Vision algorithms is identifying textured patches surrounding the keypoints across images acquired under widely varying poses and lightening conditions. Correspondences generated by the textured patches are very important for object recognition and retrieval applications. Patch recognition using correspondences is very critical step in the whole process, which both reliability and speed are important for the overall performance and usability. Achieving both reliability and speed is difficult because the correspondences can vary depending on pose changes. The standard approach to address this problem is to build affine-invariant descriptors of local image patches and to compare them across images. It results in a high computational overhead.
Random Ferns method based on ensemble classifier that relies on an offline training phase during which multiple views of the patches to be matched has been introduced to train and recognize the correspondences, using some pairwise intensity comparisons. In many object detection problems, run-time performance is very important, so Random Ferns method is more popular. It is very simple to implement, and to perform as well as SIFT while being faster. However, in pre-processing step, too much affine transform processing including Gaussian noise addition and Gaussian smoothing on each patch image has been performed for acquiring independence of pose changes. These processes need too much computation cost so it is very difficult to implement in a commercial market. Also it has another problem that affine transform significantly changes pixel data in the patch window because the pixel data before affine transform are different from those data after affine transform.
A novel technique, therefore, is proposed for addressing fast, efficient, and discriminative detection of correspondences in various imaging conditions in this thesis. Specifically, the rotation invariant HOG based descriptor is proposed for independence on pose changes. And delta-HOG is also proposed, which makes our descriptor to store the HOG in each scale layer and piles it up quickly following scale changes and considering major orientation of descriptor. In addition, the circular patch window is proposed, which can supply the descriptor consistent pixel data whenever the patch window is rotated for detecting objects under pose changes. The circular patch window uses polar coordinates. To speed up the calculation of polar coordinates in the circular patch window, every pixel location and its set of polar parameters (radius and angle from the window center) in the circular patch window is stored in a table in PC memory; therefore, it is simply read through relative addressing using the table without calculating it. The proposed method using these three kinds of methods is fast and shows good performance for the cloth pattern recognition case which is mostly under in-plain rotation and scale changes, not much in 3D pose variation. On the other hand, the entire recognition process with the proposed method using Random Ferns is proposed for deformable cloth pattern recognition and shows good performance, even when cloth patches are rotated and resized.
The proposed method is demonstrated experimentally with city, flower, and museum images where the proposed method takes 24~28 seconds for training while the previous work takes about 2 days which is impossible for internet service applications. The mismatch rate in the city, flower, and museum test images is over 50%, which is relatively very high when considering 10~20% for the previous work, but the experiments using simple images like cloth patch images show similar mismatch rate. In other words, the proposed method shows good performance in deformable cloth pattern recognition experimental results. It found a greater number of similar cloth patches than dense-SIFT in 20 tests out of a total of 36 query tests. In addition, the proposed method is much faster than dense-SIFT in both training and testing; its time consumption is decreased by 57.7% in training and 41.4% in testing, when compared with dense-SIFT on a PC with a Intel Xeon CPU 2GHz and 16GB RAM.
The proposed method, therefore, is expected to contribute to real-time cloth searching service applications that update vast numbers of cloth images posted on the internet .