CNN classification approach to motion vector detection considering video capturing satellites

Traditional block-matching methods for detecting Motion Vectors (MVs) need to calculate fitness factors for each MV detection. However, with the increasing demand for ultra-definition resolutions with higher frame rates per second, such calculations require a super processor. Artificial intelligence needs a strong processing once and, in the training, and iterations step. This article introduced a novel approach to MV detection that integrates block-matching methods with image classification via CNN. The dataset was prepared considering the limitations posed by the problem. A 17-class configuration of MVs with a 3-pixel overlap was chosen, combining pixel blocks from both frames into a single 36 × 36 pixel image for each sample. Then a CNN classification architecture was proposed, and training was performed to facilitate the recognition of motion vectors. The proposed architecture is designed to prevent overfitting. The proposed light network underwent comprehensive testing process to assess its ability to accurately identify MV PSNRs over 30 dB, obtained by comparing mainframes and reconstructed frames for four video streams captured from a satellite view, with the best being 34.86 dB, indicated that the MVs derived from this method yielded satisfactory outcomes. Additionally, the proposed architecture demonstrates more satisfactory PSNR than other methods under the same conditions. Experimental results show that can withstand up to 10% random noise in the Main Blocks. The MV directions chosen by the trained network support previous findings and suggest that the overall direction of this video is consistent with the regional direction.