International Journal of Information and Communication Technology Research (Dec 2017)
Enhancement of Illumination scheme for Adult Image Recognition
Abstract
Biometric-based techniques have emerged as the most promising option for individual recognition. This task is still a challenge for computer vision systems. Several approaches to adult image recognition, which include the deep neural network and traditional classifier, have been proposed. Different image condition factors such as expressions, occlusion, poses, and illuminations affect the facial recognition system. A reasonable amount of illumination variations between the gallery and probe images need to be taken into account in adult image recognition algorithms. In the context of adult image verification, illumination variation plays a vital role and this factor will most likely result in misclassification. Different architectures and different parameters have been tested in order to improve the classification’s accuracy. This proposed method contains four steps, which begin with Fuzzy Deep Neural Network Segmentation. This step is employed in order to segment an image based on illumination intensity. Histogram Truncation and Stretching is utilized in the second step for improving histogram distribution in the segmented area. The third step is Contrast Limited Adaptive Histogram Equalization (CLAHE). This step is used to enhance the contrast of the segmented area. Finally, DCT-II is applied and low-frequency coefficients are selected in a zigzag pattern for illumination normalization. In the proposed method, AlexNet architecture is used, which consists of 5 convolutional layers, max-pooling layers, and fully connected layers. The image is passed through a stack of convolutional layers after fuzzy neural representation, where we used filter 8 × 8. The convolutional stride is fixed to 1 pixel. After every convolution, there is a subsampling layer, which consists of a 2×2 kernel to do max pooling. This can help to reduce the training time and compute complexity of the network. The proposed scheme will be analyzed and its performance in accuracy and effectiveness will be evaluated. In this research, we have used 80,400 images, which are imported from two datasets - the Compaq and Poesia datasets - and used images found on the Internet.
