Relay Protection Setting Sheet Detection and Recognition Approach Based on YOLOv8-CRNN-CTC

Abstract

To address the challenging recognition of relay protection setting characters in industrial scenarios due to their small size, dense arrangement, complex backgrounds, and susceptibility to lighting effects, a two-stage automatic recognition method integrating an improved YOLOv8 detector and a CRNN-CTC recognizer is proposed. First, an enhanced text detection model is constructed: in the YOLOv8 backbone network, a convolutional block attention module(CBAM) is introduced to focus on text regions and suppress background interference; a weighted bi-level feature pyramid network (BiFPN) replaces the original structure to enhance multi-scale feature fusion and improve the detection capability of small characters; a lightweight detection head LWD module replaces all-scale original detection heads and uses the EIoU loss function to further optimize localization accuracy. Then, the detected text regions are cropped and input into the CRNN-CTC model for sequence recognition. This model extracts features through a convolutional neural network (CNN) and models sequence contextual information via a bidirectional long short-term memory network (Bi-LSTM), finally decoding through a connectionist temporal classification (CTC) layer to achieve recognition of variable-length character sequences. Experiments show that the text detection average precision of this method reaches 99.5%, the end-to-end character recognition accuracy exceeds 99%, significantly outperforming traditional OCR methods.