| Summary: | 碩士 === 國立雲林科技大學 === 電機工程系 === 107 === This paper develops a set of image content analysis system for mobile digital instrument. This system contains two parts: ROI detection and numerical identification. This paper uses the model-to-model method to improve the accuracy of the overall system. In the ROI detection part, this paper uses the deep learning technology SSD-MobileNet v2 to detect and identify objects in the image. Based on the pre-training model, the system can detect and identify 7 types of objects with the header category provided by the factory. Firstly, the type and position of each type of header are determined by ROI detection. After the header position is obtained, the position image is retained and enlarged to the same size as the original image, and is used as an input for numerical identification in the numerical identification part. In this paper, the object detection method is used to treat the number as an object to detect the position of the number and identify it. Through the technology of object detection, we can simultaneously turn the three steps of digital positioning, digital cutting and digital recognition into one step. The technology of object detection enhances the speed of numerical recognition and completes the instantaneous digital identification.
In order to evaluate the system performance, this paper uses Accuracy and Precision to evaluate the ROI detection and numerical identification models, and evaluates single and multiple consecutive images. In the ROI detection section, The classification of the header category of the system, each category is higher than 90% on average, and each category has an accuracy of more than 50% in numerical identification.
Keywords: object detection, ROI detection, numerical identification
|
|---|
