License Plate Recognition Model based on Improved YOLOv5 and Convolutional LSTM

An end-to-end deep learning model is proposed in this research, for licence plate recognition (LPR) and identification in natural circumstances, which addresses the accuracy and speed limitations of standard licence plate recognition approaches. By adding a better channel attention mechanism and including position data in the output, the proposed method improves the You Only Look Once (YOLOv5) down sampling process and reduces information loss during sampling for better feature extraction. An optimised the YOLO layer is used for single-class recognition to improve efficiency and accuracy. Additionally, Convolutional Long Short-Term Memory (ConvLSTM) combined with Connectionist Temporal Softmax (CTS) is used for character segmentation-free recognition. The utilization of an optimized YOLO layer for single-class recognition enhances both efficiency and accuracy. The integration of ConvLSTM in conjunction with CTS proves to be a breakthrough, facilitating faster convergence, reduced training time, and increase the precision of the model. This configuration speeds up convergence, lowers training time, and increases identification accuracy. The experimental results demonstrate average recognition precision of 99.24% and also robustness, especially in complex situations, with better performance than conventional algorithms.