Search

Popular deep learning models for text segmentation include CTPN, EAST, and PixelLink. However, they are not very well capable of dealing with the images containing densely distributed characters, and those characters may be connected. For these problems, the ResNet with excellent sensitivity for feature extraction is used to replace those embedded convolution neural networks in the main structures of CTPN and EAST. The experimental results showed that a better feature extraction network could significantly improve the precision of text localization. Noteworthy, the results indicate that the accuracy of modified EAST with ResNet101 would be the highest with a deeper depth and larger width of ResNet. The accuracy of text segmentation on ICDAR 2015 is 83.4% which is 7% higher than the original PVANET-EAST. The text detection accuracy is 83.9% on the untrained scanned document. Also, it achieved an accuracy of 86.3% when applied to self-collected Chinese calligraphy. Those results demonstrated that text detection using ResNet is a better improvement for OCR applications. [ABSTRACT FROM AUTHOR]

Most traditional facial expression-recognition systems track facial components such as eyes, eyebrows, and mouth for feature extraction. Though some of these features can provide clues for expression recognition, other finer changes of the facial muscles can also be deployed for classifying various facial expressions. This study locates facial components by active shape model to extract seven dynamic face regions (frown, nose wrinkle, two nasolabial folds, two eyebrows, and mouth). Proposed semantic facial features could then be acquired using directional gradient operators like Gabor filters and Laplacian of Gaussian. A multi-class support vector machine (SVM) was trained to classify six facial expressions (neutral, happiness, surprise, anger, disgust, and fear). The popular Cohn-Kanade database was tested and the average recognition rate reached 94.7 %. Also, 20 persons were invited for on-line test and the recognition rate was about 93 % in a real-world environment. It demonstrated that the proposed semantic facial features could effectively represent changes between facial expressions. The time complexity could be lower than the other SVM based approaches due to the less number of deployed features. [ABSTRACT FROM AUTHOR]

The general public spends a significant amount of time in front of digital signage seeking information from many venues such as exhibition halls and shopping centers. This is why advertisement purchasers believe that the number of passing viewers provides crucial information for their marketing strategies. In this paper, a real-time person counting/memorizing system is designed capable of distinguishing the gender of potential customers. An adaptive boosting (Adaboost) machine learning algorithm is used to detect human faces and utilize specific filtering criteria to eliminate useless data. For each detected person, face and torso information are recorded in a database for identification. The least recently used identification record will be deleted if the database is full. Gender classification is performed by support vector machine using hair ratios extracted from gender characterizing regions. Based on a variety of experiments, the accuracy of the proposed algorithm is higher than 90 % for person-counting and higher than 94 % for gender classification. Moreover, the execution speed on personal computers may reach 15-20 fps. [ABSTRACT FROM AUTHOR]