Boosting child abuse victim identification in Forensic Tools with hashing techniques Rubel Biswas Departamento IESA Universidad de Le´ on Researcher at INCIBE rbis@unileon.es Victor Gonz´ alez-Castro Departamento IESA Universidad de Le´ on Researcher at INCIBE vgonc@unileon.es Eduardo Fidalgo Fern´ andez Departamento IESA Universidad de Le´ on Researcher at INCIBE efidf@unileon.es Deisy Chaves Departamento IESA Universidad de Le´ on Researcher at INCIBE dchas@unileon.es Abstract—In this work, we present a scheme to identify occluded faces using perceptual image hashing. Most of the existing methods for this problem focus on occlusion detection and further removal of the occluded area by training a facial model. In this paper, we propose a new hashing method which does not require prior training. Our model combines frequency coefficients and statistical image information to increase the recognition accuracy of occluded faces. The proposed method aims to improve face recognition accuracy in forensic tools such as victim identification in Child Sexual Abuse (CSA) materials. Experimental results showed that the proposed method outper- forms the results obtained with perceptual image hashing for occluded face identification using the LFW database. Index Terms—Face identification, Face recognition, Perceptual hashing, CLOSIB, pHash, NMF Type of contribution: Ongoing research I. I NTRODUCTION Automatic face identification or recognition is widely used in many real-time applications such as forensics, surveillance or criminal identification among others. In recent years, deep learning techniques have achieved a considerable development in this area [1]. Nevertheless, there are still some open issues during face identification. One of the most challenging problems is the occlusion of the face, which can be caused by several reasons, such as self-occlusion (e.g. non-frontal posi- tion), accessories (e.g. glasses, masks or hair) or adversarial attacks (i.e. image faces modified by adding small changes to make difficult the identification). In the literature, we can find works that deal with the automatic Child Sexual Abuse (CSA) material detection [2]. However, after detecting such kind of material, Law Enforce- ment Agencies (LEA) may face the problem of identifying the victims. Children usually are presented dressed with customs, accessories that cover their faces and also sometimes the CSA material receives eye adversarial attacks before uploading the material to the Web. Therefore, occluded face identification in CSA images remains a challenging task for LEAs. To address the problem of identifying occluded faces we proposed a face recognition method based on perceptual image hashing [3] to avoid the training of facial models and represent an image content as a fixed-length vector. This research is part of the European project Forensic Against Sexual Exploitation of Children (4NSEEK), and our primary goal is to enhance the Forensic Tools ability to recognize occluded faces in child pornography materials. II. RELATED WORK Due to the availability of high-end GPU cards and training data, nowadays, deep learning methods achieve state of the art results in the task of face recognition [4]. Hongjun and Aleix [5] used SVM to find a hyperplane which is parallel to the affine subspace of occluded data. While Min et al. [6] trained a SVM classifier to detect the occluded part and use the non-occluded area to match with a corresponding part of a faces gallery. On the other hand, many perceptual image hashing methods have been proposed and applied to the field of multimedia security [7]. Researchers have focused on image hashing schemes based on the concepts of deep hashing, Discrete Co- sine Transform (DCT), Discrete Wavelet Transform (DWT), Discrete Fourier Transform (DFT), and non-negative matrix factorization (NMF), among others. We have noticed that occluded face recognition requires a trained model with occluded face features. To overcome the need for training a model, we designed an image hashing method to identify a person when the face image is partially occluded which can be applied in CSA cases. III. METHODOLOGY Fig. 1 represents the proposed perceptual hashing scheme, which is two-fold: (1) computation of the hash code of a non-occluded face image and (2) verification of the cropped occluded face. In the first stage, Multi-Task Cascade CNN (MTCNN) method [8] is applied to detect a face contained in an image. Then, the face is cropped using the detected bounding box coordinates and re-sized it to 120 ×120 pixels. Next, the perceptual hash [3] and CLOSIB descriptors [9], [10] are computed to extract 64 coefficients and 128 statistical features of the face image, respectively. Afterwards, NMF [11] is applied to reduce the 128 CLOSIB features into 64 values. Finally, an element-wise multiplication between the 64 pHash coefficients and the 64 CLOSIB features is carried out to attain the final feature vector, i.e. the hash code, for the face, which is stored. In the second stage, an occluded face is cropped manually from an image and resized into 120 ×120 pixels. After that, the image hash codes are obtained using the process previously described. Finally, the similarity score is computed between the occluded face hash code and the ones stored with a correlation coefficient function. If the score is greater than a threshold, T , it is considered that a similar face is found in the internal storage of the module.