Detection and Recognition of Professional English Texts using One Dimensional Convolutional Neural Network and Bidirectional Gated Recurrent Unit

Texts appear with many script combinations, in various orientations, sizes, colours, and backdrops. It is crucial to recognise English text accurately and quickly from a wide range of inputs to address the challenges related to complex backgrounds and handwriting variations. A two-hidden-layer 1 Dimensional Convolutional Neural Network (1DCNN) is applied to extracts features from input images. The two hidden layers consists of input network with two convolution layers, and the layer has network with Squeeze-and-Excitation (SE) block to focus on the significant features. The batch normalization to match the distribution while the max-pooling layers perform the down sampling The output of the two networks are concatenated and fed to Bidirectional Gated Recurrent Unit (Bi-GRU) network which decodes features considering of forward and backward dependencies. The last layer is the Connectionist Temporal Classification (CTC) layer which gives out the final text recognition output. Results showed that the proposed 1DCNN-Bi-GRU model showed better detection and recognition accuracy of professional English texts with 98.29%, precision of 96.34%, and recall 95.33% when compared to the existing models You Only Live Once v3 (YOLOv3) and Deep Convolutional Recurrent Neural Network (DCRNN).