FIAT/IFTA World Conference 2024

A Recurrent Convolutional Neural Network (RCNN) is being introduced to rectify inherent flaws in the SECAM format, a system extensively utilised for archiving and production from 1967 to the 1980s in many countries. Videos, preserved on archived magnetic tapes,  like two-inch tapes that encapsulate the full bandwidth of the composite signal, and U-Matic format tapes that conserve colour using a “colour under” scheme.  

These systematic artefacts, originating from past technological choices rather than ageing, were barely perceptible on 65 cm CRT TVs during their broadcast era. However, they are now deemed unacceptable for display on contemporary technology.

In the PAL format, traditional signal processing techniques enhanced the demodulation and separation of luminance and chrominance. Yet, these methods, such as linear filtering with comb filters or FFT domain filtering, are incompatible with the FM modulation of colour in SECAM. To our knowledge, no technology was previously available to enhance the quality of demodulated composite SECAM video. The proposed RCNN tackles systematic imperfections like blurring (resolution loss), mutual crosstalk between chrominance and luminance, 12.5 Hz colour flicker, and chroma and luma noise. The flicker of horizontal lines is a consequence of the unfiltered vertical sampling of chroma in the SECAM standard.

To train the RCNN, a degraded dataset was generated by processing recent videos through a physical SECAM modulator/demodulator chain. The RCNN also transforms interlaced frames into progressive images, thereby enhancing visual quality and rendering them more suitable for further processing by other super-resolution AI techniques. The processing is executed at a rate of several frames per second on a recent GPU on the already demodulated image YDrDb. This enables the treatment of programmes that have already been digitised and compressed, without requiring access to the original composite video.