Method of Eliminating Reflection Interference for Image Recognition in Autonomous Driving

In the field of automatic driving and visual detection, it is easy to be disturbed by the reflective light and the mirror surface of the reflective images. To overcome it, this paper proposes a mirror recognition model using EfficientNetV2 Medium as the backbone module. And we have utilized parallel convolutional layers and a lightweight convolutional attention module based on guided convolutional neural networks to capture low-level and high-level features for edge detection. The research results indicate that the model has better recognition accuracy, recall and F1 value, and the accuracy of 0.8331, maximum recall of 0.8366 and maximum F1 value in the test set are 0.8211. The research results provide a theoretical method for eliminating mirror image interference, which can be applied to autonomous driving and assisted safe driving.