Container keyhole positioning based on deep neural network

In recent years, more and more automated container ports have increased the requirements for the accuracy and real-time performance of container keyhole identification. In this paper, the improved deep neural network algorithm YOLO is used to identify the position of the keyhole. Compared with the original method, this method increases the dimensionality reduction of the input vector and the accurate extraction of the subsequent target area, shortens the detection time and improves the accuracy. The model trained in this paper has the mean Average Precision (mAP) of 87.7% under the test set, the accuracy rate of 96%, the recall rate of 83%, and an Intersection-Over-Union (IOU) of 80.43%. The detection time of an image on GPU is 10 ms, the detection time on CPU is 80 ms, and the frame rate of the actual detected video reaches 15FPS. This study provides a theoretical basis for automatic positioning of container keyholes.