Research On Robot Object Recognition Method Based On Scene Text Reading

Robot object recognition is one of key technologies in the development of service robot.Accurately perceiving and recognizing the scene object in complex scenes is a challenge of robotics technology.Scene text reading can deepen the robot's object recognition ability,better path planning,and real-time visual translation.The rapid development of robotics and the arrival of the aging society increase the significant demand for home service robots.Fetching is one of the general tasks.There are many kinds of objects in the home environment,and high accuracy is essential.For example,the appearance of medicines is mostly similar,and wrong recognition will lead to serious consequences.So robot should have strong object recognition ability.At present,the object recognition method based on deep learning mainly identifies object shapes.Train with large different types of object images,and then classification recognition.However,fetching household objects is challenging because of the demand of recognizing different shape objects in the same category,or similar shapes from different categories.Therefore,recognizing objects only by shape is insufficient.With the aim to solve issues of robot object recognition in complex scenes,this paper proposes an object recognition method based on scene text reading.The proposed method simulates human-like behavior and accurately identifies objects with texts through careful reading.The robot detects and recognizes texts in the image and then stores the recognition results in a text file.When the user gives the robot a fetching instruction,the robot searches corresponding keywords from the text files and achieves the confidence of multiple objects in the scene image.Then,the object with the maximum confidence is the target.The results show that the robot can accurately distinguish objects with arbitrary shape and category,and effectively solve the problem of object recognition in home environments.The main contributions of this paper include the following three aspects:(1)This paper proposes an object recognition approach based on scene text reading.Innovatively combines text recognition with robot object recognition.With this approach,robots can recognize arbitrary-shape and arbitrary-category objects.In addition,this approach speeds up robot mimicking of human recognition behavior.(2)To improve the recognition accuracy of models,we generate a new dataset and its inverse.The generated dataset contains 102,000 images with labeled documents while the inverse dataset inverts the pixel value of generated images without changing the labels.After training on these datasets,the recognition accuracy of the model is improved by 1.26%.(3)Experiments are carried out on relations between confidence thresholds of text boxes and recognition effect.The confidence threshold is higher and the recognition results are more accurate.However,useful information may be missed.By statistics of test samples,the confidence threshold is set at 0.97,a good balance that indicates the key information is reserved and the recognition accuracy is high with few wrong words.