Improved Neural Network using Integral-RELU based Prevention Activation for Face Detection

Numerous variation of neural networks improved the face detection performance significantly. However, extravagant computation becomes a major problem found when the window explored the non-face candidate. To prevent useless detection in the non-face area, our proposed method is the addition of a prevention function using integral representation in the RELU (Rectified Linear Units) activation function. If the integral of RELU does not reach the threshold, the convolution is skipped and the window shifts to the neighboring area. Both functions are selected because they are easy to calculate and achieve convergence speed rapidly. Based on the results of trials on 10 data, the Integral CNN + Integral is faster than CNN + RELU with a speed ratio of 125: 285 FPS. Besides, ‘RELU CNN + Integral’ has fewer face detection redundancies compared to Viola & Jones algorithms. This result shows that our proposed method is superior to the state-of-the-art algorithms.