Key-point based license plate detection using fully convolutional neural networks

License plate detection plays an important role in automatic recognition of license plate numbers. Region-based detection for this task, however, is hampered by the long processing time and the less robustness to appearance variations. To address these issues, key-point search has become a promising approach since key-points are highly invariant in the sense that their appearances are not drastically affected by changes in environmental conditions. This research proposed modeling of key-points in the form of heat-maps such that key-point detection can be conducted using the indirect method of convolutional regression. Convolutional regression has the ability to preserve spatial information leading to consistent and accurate detection. A convolutional neural network (CNN) that comprises convolutional layers and de-convolutional layers is constructed in this research to perform the regression. This CNN produces five heat-maps to detect five types of key-points namely four corner points and a centroid. Using these five key-points, an efficient algorithm to determine bounding polygons of license plates is also formulated. The proposed method is evaluated on video frames collected from a surveillance camera installed at an entrance gate. The experimental results demonstrate that the method is able to achieve an average precision of 93.8% with a processing time of 0.055 seconds/frame. These results are significantly better than those achieved by the Faster R-CNN method that achieves an average precision of 77.7% with a processing time of 3.198 seconds/frame on the same dataset.