| abstract |
An online detection and recognition method for deep-sea targets with multi-threshold dynamic statistics belongs to the field of machine vision. The invention makes use of the underwater robot platform to fully consider the environmental information of the deep-sea area, and corrects the type and number of rectangles generated by anchor points in the convolutional neural network; when training the model, the minimum stable numerical value detection program is added to the loss function module, and the minimum stable numerical value detection program is added in the stable minimum Stop the training at the value to get the best model; change the number ratio of the training, verification and detection pictures in the produced data set to get different training results, select the model with the highest detection and recognition accuracy to add to the underwater Real-time target detection and recognition are carried out in the vision module of the robot; an improved multi-threshold dynamic windowing method is used to detect the number of targets, and the detection zone of each threshold height detects a different number of targets for counting and statistics. Compared with the traditional target detection and identification method, the invention is faster and more accurate, and has stronger adaptability to complex environment. |