Some examples of face and license plates are shown in Figure 6. The background Regorafenib clinical trial images were collected by randomly selecting 150,000 frames from an arbitrary image set that does not contain any face or any license plate. The size of a background image is the same as the face images or the license plate images. Since object (face or license plate) detection test needs far more background images as negative samples, additional images were generated by the same method.Figure 6Training image examples of faces and license plates.4.2. Cross-Validation TestTo evaluate the accuracy performance of the classifier, a 5-fold cross-validation test in terms of error rate was performed with 100,000 positive samples (faces, license plates) and 150,000 negative samples (backgrounds).

All samples were divided into five groups that have 20,000 positive samples and 30,000 negative samples each. A single group was retained as the validation set for testing the trained classifier, and the remaining four groups were used as the training set. Thus, the cross validation process can be repeated five times with each validation set. The five results of error rates can be averaged to produce the overall performance.Figures Figures77 and and88 show the comparison of the performances according to the number of pixel locations combined and used in the strong classifier. The numbers of pixels that were compared were 20, 40, 80, 160, and 320. Horizontal axis marks the number of rounds for Adaboost training while vertical axis shows the error rate. The total number of rounds for Adaboost training was 1000.

Figure 7 shows the results of training error. Red lines represent the results on face DB, and blue lines represent the results on license plate DB. Dotted lines and solid lines represent the results when the numbers of combined pixels are 20 and 40, respectively. As shown in Figure 7, the training error rate for the face DB converges to zero with 20 pixels at round 37 and 40 pixels at round 31. The training error rate for the license plate DB converges to zero with 20 pixels at round 57 and 40 pixels at round 38. The training error rate of 80, 160, or 320 is the same as the training error rate of 40. Thus, the training error rate converges to zero rapidly when the number of combined pixels becomes larger. Figure 8 shows the results on testing error for face DB and license plate DB, respectively.

Through the experimental results, the testing error rate is closed to zero when the number of combined pixels becomes larger at the same round. For the number of pixels above 40, it is observed that the testing error Batimastat rate results diverged, although the results of training error rate were uniform. It is noted that when the number of combined pixels is too low to construct the strong classifier, such as for 20 pixels, the testing error rate decreases up to round 40 but increases beyond round 40.