Design of a Light-weight Natural Scene Text Detection System

Scene Text Detection is one of the vast tasks of computer vision that have received immense progress in the field of deep learning (DL). Several DL models have been made with large scale neural network structures and heavy weight files with a huge number of parameters to increase the natural scene text detection performance. Although these models improve the performance to a high range, but they reduces the speed of the model and it also becomes difficult to use them in platforms with limited resources. Therefore, the goal of our proposed work lies in making models smaller, faster and more efficient. Our proposed model is made in the form of encoder-decoder architecture where the light-weight model, MobileNetV2 is used as the encoder. The decoder is made by concatenating certain feature maps from the encoder with the transposed convolutional layers. After training, the proposed model generates a weight file of 4 MB with 1.1 M parameters. The proposed model can be trained in Google Colaboratory with 12 GB RAM and without using GPU in few hours which shows the speed of the model and which gives the advantage of making the proposed work at free of cost. The performance of the proposed model on Total-Text and ICDAR 2013 datasets shows that the proposed work achieves a desirable performance in detecting texts from natural scene images.