Text detection for dust image based on deep learning

With a large number of images, we rely on the text in images to understand the image quickly and extract useful information. So text detection in images has become one of the important topics of intelligent information processing. With the deteriorating environment, dust weather is more and more common, and text regions in dust images always exit some questions that blur, text features weakened or lost greatly affects our understanding of images and also limits extracting text information from images. Therefore, the traditional algorithm is not good at text detection of dust images. In order to solve these problems, a new method is proposed to detect the dust image text, which is divided into two modules. Firstly, the dust image is enhanced by the dust image enhancement algorithm based on color transfer, and the text and non-text regions in the image are divided by the maximally stable extremal regions, which greatly reduces the computational cost. Next, we choose the text candidate region through convolutional neural network. And then text lines are obtained by the run length smoothing algorithm. Finally, the non-text regions are removed by gaussian smoothing and candidate area filtering to realize text detection in dust images. The experimental results show that the algorithm can be used to detect text regions in dust images with good performance.