However, it seems that overfitting does not occur when the number of combined pixels becomes greater than 80.Figure 7Training error according to the number of pixel locations.Figure 8(a) Testing error according to the number of pixel locations on face DB. (b) Testing error according to the number of pixel www.selleckchem.com/products/nutlin-3a.html locations on license plate (LP) DB.Table 1 shows the execution time of classification and the number of misclassified samples for 200,000 test samples on a MATLAB platform. The classifier used to test was trained with 20,000 positive samples and 30,000 negative samples that were not included in the test samples. For the assessment to be reliable, the experiments for computing execution times were repeated ten times and the total execution time and their averages were calculated.

Each column of the table represents the number of combined pixels used for the strong classifier and all the tested classifiers were trained through the same 1000 rounds. When the number of pixels decreases, the processing speed becomes faster while the error rate increases. On the contrary, when the number of pixels increases, the processing speed becomes slower and the error rate decreases. Therefore, in order to configure a cascade classifier as shown in [17�C19], the strong classifier consisting of a small number of pixels was placed at earlier stage and a large number of pixels at later stage.Table 1Execution time of classification and number of misclassified samples for 200,000 test samples.Figure 9 shows a comparison of learning algorithms composed of the proposed algorithm and the conventional algorithm used in [17�C19, 25].

The number of combined pixels for the strong classifier was 320. Red lines and blue lines represent the results on face DB and license plate DB, respectively. Dotted lines illustrate the results of the conventional algorithm, and solid lines depict the results of the proposed algorithm. For the training error rate as shown in Figure 9(a), the proposed algorithm has better performance since its training error rate converges to zero quicker with both databases. According to the results, the training error rate for the face DB converges to zero with the proposed algorithm at round 31 and the conventional algorithm at round 48. The training error rate for the license plate DB converges to zero with the proposed algorithm at round 38 and the conventional algorithm at round 61.

Figure 9(b) shows a comparison of performance of the proposed and conventional algorithms for testing error. Through the experimental results, it can be claimed that the error rate of the proposed algorithm converged to zero faster than that of the conventional algorithm. Carfilzomib Therefore, the classifier based on the proposed algorithm exhibited better performance with the same number of rounds compared to the classifier based on the conventional algorithm.