Marco Fontani

The need to estimate the speed of road vehicles has become increasingly important in the field of video forensics, particularly with the widespread deployment of CCTV cameras worldwide. Despite the development of various approaches, the accuracy of forensic speed estimation from real-world footage remains highly dependent on several factors, including camera specifications, acquisition methods, spatial and temporal resolution, compression methods, and scene perspective, all of which can significantly influence performance. In this paper, we introduce ForeSpeed, a comprehensive dataset designed to support the evaluation of speed estimation techniques in real-world scenarios using CCTV footage. The dataset includes recordings of 14 vehicle passes at known speeds, captured by three digital and three analog cameras from two distinct perspectives. Real-world road metrics are provided to enable the restoration of the scene geometry. Videos were exported with multiple compression levels and settings to simulate real-world scenarios in which export procedures are not always performed according to forensic standards. Overall, ForeSpeed includes a collection of 322 videos. As a case study, we employed the ForeSpeed dataset to benchmark a speed estimation algorithm available in a commercial product (Amped FIVE). The results demonstrate that while the method reliably estimates average speed across various conditions, the coverage rate and the estimated error range can be significantly affected by perspective. Furthermore, having higher spatial or temporal resolution alone does not guarantee better reliability of the estimates. The ForeSpeed dataset is publicly available to the forensic community, with the aim of facilitating the evaluation of current methodologies and inspiring the development of new, robust solutions tailored to collision investigation and forensic incident analysis.

This study proposes a method to detect double compression in H.265/HEVC videos containing B-frames, a scenario underexplored in previous research. The method extracts frame-level encoding features—including frame type, coding unit (CU) size, quantization parameter (QP), and prediction modes—and represents each video as a 28-dimensional feature vector. A bidirectional Long Short-Term Memory (Bi-LSTM) classifier is then trained to model temporal inconsistencies introduced during recompression. To evaluate the method, we created a dataset of 129 HEVC-encoded YUV videos derived from 43 original sequences, covering various bitrate combinations and GOP structures. The proposed method achieved a detection accuracy of 80.06%, outperforming two existing baselines. These results demonstrate the practical applicability of the proposed approach in realistic double compression scenarios.

Extensive research has been conducted in image forensics on the analysis of double-compressed images, particularly in the widely adopted JPEG format. However, there is a lack of methods to detect double compression in the HEIF format, which has recently gained popularity since it allows for reduced file size while maintaining image quality. Traditional JPEG-based techniques do not apply to HEIF due to its distinct encoding algorithms. We previously proposed a method to detect double compression in HEIF images based on Farid’s work on coding ghosts in JPEG images. However, this method was limited to scenarios where the quality parameter used for the first encoding was larger than for the second encoding. In this study, we propose a lightweight image classifier to extend the existing model, enabling the identification of double-compressed images without heavily depending on the input image’s quantization history. This extended model outperforms the previous approach and, despite its lightness, demonstrates excellent detection accuracy.

The last years have witnessed significant developments in image acquisition systems and in algorithms for extracting information from them. Nevertheless, in many scenarios, several factors can hinder the recovery of useful data from images. This is especially true and important in forensic applications, where images are often accidentally captured by an imaging system not engineered for that specific acquisition (for example, the surveillance system designed to monitor the entrance of a bank may accidentally capture the license plate of a vehicle passing outside the bank). Therefore, the acquired images often need to be processed to facilitate extracting information from them. When facing a combination of several impairment factors, such as blur and perspective distortion, several image restoration algorithms must be applied. Then, it is necessary to choose a restoration order, which means the order by which single restoration algorithms are chained together to obtain the enhanced image. This study aims to understand whether such an order may impact the final result. Of course, there exists a wide variety of image impairments; in this study, we focus on the case of an image affected by a combination of optical/motion blur, perspective distortion, and additive noise, which are all widespread artifacts in forensic image applications. To answer the question about the importance of choosing one restoration workflow over another, we first model each considered defect and its restoration operator and then analyze and compare the effects of the composition of such operators on the restored output. Such a comparison is made from both a mathematical and experimental point of view, using both images with synthetically generated impairments and pictures with real degradations. The results show that the restoration order can affect significantly the results, especially when the defects are severe.

Multimedia data manipulation and forgery has never been easier than today, thanks to the power of Artificial Intelligence (AI). AI-generated fake content, commonly called Deepfakes, have been raising new issues and concerns, but also new challenges for the research community. The Deepfake detection task has become widely addressed, but unfortunately, approaches in the literature suffer from generalization issues. In this paper, the Face Deepfake Detection and Reconstruction Challenge is described. Two different tasks were proposed to the participants: (i) creating a Deepfake detector capable of working in an “in the wild” scenario; (ii) creating a method capable of reconstructing original images from Deepfakes. Real images from CelebA and FFHQ and Deepfake images created by StarGAN, StarGAN-v2, StyleGAN, StyleGAN2, AttGAN and GDWCT were collected for the competition. The winning teams were chosen with respect to the highest classification accuracy value (Task I) and “minimum average distance to Manhattan” (Task II). Deep Learning algorithms, particularly those based on the EfficientNet architecture, achieved the best results in Task I. No winners were proclaimed for Task II. A detailed discussion of teams’ proposed methods with corresponding ranking is presented in this paper.

The last decade witnessed a renaissance of machine learning for image processing. Super-resolution (SR) is one of the areas where deep learning techniques have achieved impressive results, with a specific focus on the SR of facial images. Examining and comparing facial images is one of the critical activities in forensic video analysis; a compelling question is thus whether recent SR techniques could help face recognition (FR) made by a human operator, especially in the challenging scenario where very low resolution images are available, which is typical of surveillance recordings. This paper addresses such a question through a simple yet insightful experiment: we used two state-of-the-art deep learning-based SR algorithms to enhance some very low-resolution faces of 30 worldwide celebrities. We then asked a heterogeneous group of more than 130 individuals to recognize them and compared the recognition accuracy against the one achieved by presenting a simple bicubic-interpolated version of the same faces. Results are somehow surprising: despite an undisputed general superiority of SR-enhanced images in terms of visual appearance, SR techniques brought no considerable advantage in overall recognition accuracy.

Photo Response Non-Uniformity (PRNU) is considered the most effective trace for the image source attribution task. Its uniqueness ensures that the sensor pattern noises extracted from different cameras are strongly uncorrelated, even when they belong to the same camera model. However, with the advent of computational photography, most recent devices heavily process the acquired pixels, possibly introducing non-unique artifacts that may reduce PRNU noise's distinctiveness, especially when several exemplars of the same device model are involved in the analysis. Considering that PRNU is an image forensic technology that finds actual and wide use by law enforcement agencies worldwide, it is essential to keep validating such technology on recent devices as they appear. In this paper, we perform an extensive testing campaign on over 33.000 Flickr images belonging to 45 smartphone and 25 DSLR camera models released recently to determine how widespread the issue is and which is the plausible cause. Experiments highlight that most brands, like Samsung, Huawei, Canon, Nikon, Fujifilm, Sigma, and Leica, are strongly affected by this issue. We show that the primary cause of high false alarm rates cannot be directly related to specific camera models, firmware, nor image contents. It is evident that the effectiveness of PRNU based source identification on the most recent devices must be reconsidered in light of these results. Therefore, this paper is intended as a call to action for the scientific community rather than a complete treatment of the subject. Moreover, we believe publishing these data is important to raise awareness about a possible issue with PRNU reliability in the law enforcement world.

The Variation of Prediction Footprint (VPF), formerly used in video forensics for double compression detection and GOP size estimation, is comprehensively investigated to improve its acquisition capabilities and extend its use to video sequences that contain bi-directional frames (B-frames). By relying on a universal rate-distortion analysis applied to a generic double compression scheme, we first explain the rationale behind the presence of the VPF in double compressed videos and then justify the need of exploiting a new source of information such as the motion vectors, to enhance the VPF acquisition process. Finally, we describe the shifted VPF induced by the presence of B-frames and detail how to compensate the shift to avoid misguided GOP size estimations. The experimental results show that the proposed Generalized VPF (G-VPF) technique outperforms the state of the art, not only in terms of double compression detection and GOP size estimation, but also in reducing computational time.