Natural Scene Text Detection Algorithm Based on Improved DBNet

There are many problems in text detection, such as large scale differences of high-resolution image features and poor multi-scale feature fusion, we propose an improved algorithm based on dbnet. On the basis of the feature fusion module, we add a atrous Convolution network with kernel-shared pooling to increase the receptive field, so that higher-level semantic information can be obtained in the feature fusion network, and through the shared kernel, the number of model parameters can be reduced, the computational cost can be reduced, and the detection accuracy can be improved. At the same time, we add the attention mechanism into the residual network to suppress the complex background noise and promote the information interaction between channels. In the loss function, we use dice loss partially to solve the imbalance of positive and negative sample data. Our experimental evaluation is on ICDAR2013 and ICDAR2015 datasets. The experimental results show that the algorithm has a certain improvement in accuracy and F value.