Double-stage article detection method and device

Abstract

The invention discloses a double-stage article detection method and device, and the method comprises the steps: inputting a real-time image into a foreground detection model in response to the obtained and collected real-time image, obtaining foreground positioning information, and determining the position of at least one prediction frame containing an article in the real-time image based on the foreground positioning information; intercepting the real-time image according to the position of the at least one prediction frame to obtain at least one prediction frame image, inputting the at least one prediction frame image into a category judgment model, and outputting the category of the at least one article; and in response to the obtained foreground positioning information, obtaining position information of the at least one article based on binocular image parallax calculation. By dividing article detection into two processes of foreground analysis and category judgment, the problem that a classical deep detection algorithm is difficult to collect enough samples to complete training in application is effectively avoided, and the contradiction between the portability of a blind guiding device and the calculation complexity of a deep learning model is effectively solved.

Description

technical field [0001] The invention belongs to the technical field of object detection, and in particular relates to a two-stage object detection method and device. Background technique [0002] Data show that with the increase of population and the deepening of aging, by 2050, it is estimated that 703 million people in the world will face moderate to severe visual impairment or blindness. According to the data of the China Disabled Persons' Federation, there are at least 5 million blind people in my country at present, and the number of blind people is rapidly increasing at a rate of 450,000 per year. Vision is the most important means of perception for human beings, and about 90% of human perception information comes from the eyes. Because of the lack of means of visual perception, the life of blind people is extremely inconvenient, and it also brings a heavy burden to society. How to enhance the autonomous environment perception ability of blind people has always been ...

AI Technical Summary

Problems solved by technology

For this reason, the present invention provides a two-stage object detection method, which considers that the traditional detection method uses a single neural network model to obtain the position and category of the object to be detected at one time, the training is difficult, and the problem of over-fitting is easy under small sample conditions. The position detection and category discrimination are divided into two models for training separately. The foreground detection model only trains possible item location information, and the category discrimination model is only used to classify possible items. Since each model only completes part of the detection process, it can It effectively reduces the difficulty of model training, and at the same time adopts the transfer learning method in the model training process. The training of the two sub-models is carried out on the basis of the mature thousand-category network model parameters, and the feature extraction function of the thousand-category network is retained as much as possible, ensuring The effectiveness of feature extraction; in addition, by selecting an appropriate category discriminant network model, the misrecognition rate of untrained items is reduced, thereby improving the success rate of the detection model under the condition of small sample set training

Images