Natacha Ruchaud

Learned vs. Hand-Crafted Features for Pedestrian Gender Recognition

This paper addresses the problem of image features selection for pedestrian gender recognition. Hand-crafted features (such as HOG) are compared with learned features which are obtained by training convolutional neural networks. The comparison is performed on the recently created collection of versatile pedestrian datasets which allows us to evaluate the impact of dataset properties on the performance of features. The study shows that hand-crafted and learned features perform equally well on small-sized homogeneous datasets. However, learned features significantly outperform hand-crafted ones in the case of heterogeneous and unfamiliar (unseen) datasets. Our best model which is based on learned features obtains 79% average recognition rate on completely unseen datasets. We also show that a relatively small convolutional neural network is able to produce competitive features even with little training data.

Privacy protecting, intelligibility preserving video surveillance

Video surveillance is increasingly omni-present in our everyday life and is a key component of many security systems. Not only is the increasing number of cameras, but also the resolution of visual sensors and the performance of video processing algorithms. This evolution generates some important privacy concerns. This article introduces a new visual filter that includes a good trade-off between privacy and intelligibility. It ensures that people are unrecognizable while keeping the scene understandable in terms of events which allows machines to detect abnormal behavior. The algorithm operates in the DCT domain to be compliant with the popular JPEG and MPEG codecs. For each sensitive area of the picture (i.e. area where privacy needs to be protected), the proposed algorithm uses the low-frequency coefficients of the DCT to display a privacy preserved image of the region and the high-frequency coefficients to hide most of the original information. Finally, our process allows authorized users to nearly reverse the process thanks to the hidden information.

The impact of privacy protection filters on gender recognition

Deep learning-based algorithms have become increasingly efficient in recognition and detection tasks, especially when they are trained on large-scale datasets. Such recent success has led to a speculation that deep learning methods are comparable to or even outperform human visual system in its ability to detect and recognize objects and their features. In this paper, we focus on the specific task of gender recognition in images when they have been processed by privacy protection filters (e.g., blurring, masking, and pixelization) applied at different strengths. Assuming a privacy protection scenario, we compare the performance of state of the art deep learning algorithms with a subjective evaluation obtained via crowdsourcing to understand how privacy protection filters affect both machine and human vision.

JPEG-based scalable privacy protection and image data utility preservation

Here, the authors propose a scalable scrambling algorithm operating in the discrete cosine transform (DCT) domain within the JPEG codec. The goal is to ensure that people are no more identifiable while keeping their actions still understandable regardless of the image size. For each 8 × 8 block, the authors encrypt the DCT coefficients to protect data information, and shift them towards the high frequencies to make the DC position available. Whereas encrypted coefficients appear as noise in the protected image, the DC position is dedicated to restitute some of the original information (e.g. the average colour associated with one or a group of blocks). The proposed approach automatically sets the value of each DC according to the region of interest size in order to keep the level of privacy protection strong enough. Comparing to existing methods, the proposed privacy protection framework provides flexibility concerning the appearance of the protected version which makes it stronger for protecting the privacy even during potential attacks. Moreover, the method does not cause excessive perturbation for the recognition of the actions and slightly decreases the efficiency of the JPEG standard.

ASePPI, an adaptive scrambling enabling privacy protection and intelligibility in H.264/AVC

The usage of video surveillance systems increases more and more every year and protecting people privacy becomes a serious concern. In this paper, we present ASePPI, an Adaptive Scrambling enabling Privacy Protection and Intelligibility. It operates in the DCT domain within the H.264 standard. For each residual block of the luminance channel inside the region of interest, we encrypt the coefficients. Whereas encrypted coefficients appear as noise in the protected image, the DC value is dedicated to restore some of the original information. Thus, the proposed approach automatically adapts the level of protection according to the resolution of the region of interest. Comparing to existing methods, our framework provides better privacy protection with some flexibilities on the appearance of the protected version yielding better visibility of the scene for monitoring. Moreover, the impact on the source coding stream is negligible. Indeed, the results demonstrate a slight decrease in the quality of the reconstructed images and a small percentage of bits overhead.

ASePPI: Robust Privacy Protection Against De-Anonymization Attacks

The evolution of the video surveillance systems generates questions concerning protection of individual privacy. In this paper, we design ASePPI, an Adaptive Scrambling enabling Privacy Protection and Intelligibility method operating in the H.264/AVC stream with the aim to be robust against de-anonymization attacks targeting the restoration of the original image and the re-identification of people. The proposed approach automatically adapts the level of protection according to the resolution of the region of interest. Compared to existing methods, our framework provides a better trade-off between the privacy protection and the visibility of the scene with robustness against de-anonymization attacks. Moreover, the impact on the source coding stream is negligible.

De-genderization by body contours reshaping

This paper deals with privacy protection in video surveillance. More specifically, the main goal of this work is to make the gender of people no more recognizable while preserving enough information concerning body shape and motion of people for action classification. We denote this processing as de-genderization. Regarding the current state-of-art methods, most of them have privacy filters only dedicated to de-identify people. These methods do not automatically imply the suppression of visual semantic traits such as gender. Therefore, we propose two approaches that modify the visual appearance of the body shape in order to de-genderize people while keeping the possibility to interpret the video. In both methods we start by extracting the contour points attached to the body shape of people in videos. Then we either mix the coordinates of the body shape and a predefined model, or we smooth the body shape by successive polygonal approximations based on convexity. Our results demonstrate that both proposed approaches protect the gender information while preserving the global body movement. The second approach based on convexity better preserves the visibility of human activities.