||Face detection is mainly used to determinate wherther the face in the image can be detected or not. Accordingly the identity of the person can be recognized. In IoT (Internet of Things) applications, face detection plays an important role, such as identity recognition for door-accessing in buildings, and customs. However, high temperature, power switch, data transmission, defects, and wear-out of image processing circuits of the face detection system may result in noisy or erroneous images. Fortunately, these noises or errors do not necessarily fail the face detection system. For example, if the structure of the face is not seriously destroyed, the face is still likely detectable. In other words, the face detection system contains the error-tolerance feature where some errors can be accepted. The lifetime of the system can thus be extended.|
In the thesis, we carefully analyze the error tolerability of face detection systems, and compare with the human visiual system. Interestingly the comparison results show that the face detection system even has much higher error tolerability. We also propose a high efficiency error-tolerance test method to examine the acceptability of an image. In our experiments a total of 5730 erroneous benchmark images are used to evaluate the test accuracy of the proposed test method. Experimental results show that 97.12% test accuracy is achieved.
In addition, in this thesis we also present an efficient image repair method for face detection systems. The results show that the proposed method can effectively enhance the acceptability of erroneous images, and thus improve the reliability of the face detection system. In the literature there have been a number of related repair works. However most of these researches focus on repairing noisy images. Erroenous images due to circuit aging are seldom considered. Although some work targets such images, single-bit errors are assumed. In the thesis, we consider not only erroneous images, but also the occurrence of multiple-bit errors. We also investigate the issues that the previous approaches cannot deal with such errors effectively. Accordingly we propose a more powerful method to repair erroneous image. It is worth mentioning that our developed method has excellent adaptability where our method can be dynamically reconfigured according to the significance of erroneous images. As a result, the content of high-quality images can still be maintained, while the quality of the moderate-quality images can be significantly increased. As for low-quality images, they will not be repaired in order to save computation time and power consumption.