A Modification of the Temporal Group AttentioMethod on Super-Resolution Video for Vehicle Number Plate Detection

Today, the development of smart cities is relatively rapid, and one of the components of a smart city is the Intelligent Transportation System (ITS). Vehicle detection and assistance on the road is part of ITS, which can be done by identifying the vehicle’s license plate. In the following study, the authors conducted the detection and assistance of motorized vehicle license plates by modifying the TGA network. Altering the TGA network on super-resolution video increases the accuracy and speed of detecting vehicle license plates. Furthermore, with network modification, the temporal information retrieved can be more balanced and have a more dynamic structure. We made modifications in several parts: (i) using the PDC (Pyramid Deformable Convolution) method for registration, where this process did not exist in the previous TGA. (ii) on Temporal Grouping, group division is made dynamic depending on the number of input frames. (iii) on the Intra Group Fusion Module, using the MSTC (Mixed Spatial-Temporal Convolution) method, while on the previous TGA, using 3D convolution. (iv) reducing the number of repetitions of the method DSC (Dense Skip Connections) in feature extraction and reducing the use of GCSC (Group Convolution with Skip Connections) only once in the Intra Group Fusion Module process. According to the experiment, our modification can boost accuracy by up to 4.23% while maintaining the same computation time. Therefore, this research contributes positively to the problem of vehicle license plate detection.