DEEP-FIR

Several countries around the world use CCTV systems as forensic evidence to combat crime. These cameras cover large fields of view, where low-resolution facial images are typically captured, making the identification of the subject of interest very difficult. Moreover, distortions caused by video compression, motion blur, and poor lighting conditions can further reduce quality and thus reducing their effectiveness. Some commercial products have recently included super-resolution techniques that fuse consecutive video frames to restore higher quality images. Nevertheless, these methods are in most cases insufficient, especially when dealing with dynamic non-rigid objects such as faces. The problem addressed by this project is to improve the quality of facial images captured by CCTV cameras using models optimized to restore compressed low-resolution facial images typically found in CCTV footages. The primary investigator has developed an algorithm able to restore low-quality facial images using artificial intelligence (AI) techniques. Extensive experiments using more than 8,000 images conducted in a relevant environment show significant gains in terms of both quality and recognition (between 20-30% improvements over state-of-the-art). However, this method is limited to restore only the facial region, uses a sub-optimal process to select the dominant feature-vectors and is unable to restore the nonlinear artefacts caused by compression. The aim of the proposed product is to enhance this method using more advanced AI techniques, with the following advantages:
 – The method will learn the filters that minimize distortion from the training data without the need to identify the dominant feature-vectors.
 – Users will only have to select the face to restore which reduces the manual labour.
 – It will be able to restore the whole head, including the hair region important for person identification.
 – It will be able to restore extensively compressed images, which is usually the case for CCTV using models that are robust to nonlinearities.
 – The developed method will provide reproducible results.
 – Reduce the computational complexity of the algorithm.
 – It restores facial images with higher quality than existing forensic tools.
 The developed algorithm will be tested on real-world CCTV videos and compared against existing video forensic tools used by forensic experts in their labs. Apart from video forensics, the proposed technology can be adopted and used in other sectors such as the video analytics and iris recognition, where we have already attained positive preliminary results."