Neural-network-based Approach to Detect and Recognize Distorted Text in Images with Complicated Background

To address the unsatisfactory recognition of distorted text in images with complicated background, a neural-network-based approach was proposed to detect and recognize text. For detection of distorted text, the improved CRAFT model was applied, and deformable convolution was introduced to replace conventional convolution to sufficiently extract the features with irregular background. On this basis, the CRAFT-DCN text detection model was proposed to improve the accuracy of text detection. In order to reduce the interference of distorted text on the recognition model, images of separated texts were tailored according to the coordinates obtained by the detection model. Meanwhile, the Dense-CRNN model was designed, and the dense convolutional layer was introduced in the text recognition model to enhance the reuse of the features, thereby reducing interference of complicated background and recognizing separated distorted text correctly. The experiment results show that, compared with traditional approaches, the improved method introduced in this paper has better detection and recognition rates. And specifically, its text detection accuracy and the text recognition accuracy in actual scenario reach 86.3% and 95.3% respectively.