Improved Angle Steel Character Detection Algorithm Based on DBNet

Abstract

Aiming at the problems of false detection and missed detection in manual recording of industrial angle steel characters, an improved angle steel character detection algorithm based on DBNet is proposed. First, the angle steel character dataset is collected. The algorithm uses MobileNetV3 as the backbone network, and replaces the attention mechanism in the backbone network with a new coordinate attention mechanism to obtain larger regional information and enhance the ability to extract complex features; secondly, weighted bidirectional The feature pyramid network improves the efficiency of multi-scale fusion and enhances the detection ability of small features; finally, a new data augmentation method and learning rate optimization strategy are used to enhance the robustness and convergence accuracy of the model. The experimental results show that, compared with the original DBNet algorithm, the detection index of the improved DBNet algorithm is increased from 81.42% to 99.06% on the angle steel character data set, which meets the needs of angle steel character detection in the industrial environment.