Research on Raindrop Image Segmentation Method Based on Improved BiSeNet

With the development of intelligent vehicles, to reduce the manufacturing cost of the whole car and improve the intellectual level of the entire vehicle, the raindrop image on the windshield is collected by using the camera in the car. Then the raindrop is automatically identified in combination with the intelligent algorithm, replacing the traditional infrared equipment in the raindrop identification to save cost and improve automation and accuracy. Based on the raindrop image captured by the dashcam, this paper improves the general semantic segmentation network BiSeNet V2, introduces the structure of the lightweight network MobileNet V3 to reduce the number of model parameters and calculation, and adds the spatial attention mechanism and Squeeze-and-Excitation Network to improve the segmentation accuracy further and realize the accurate identification of raindrops; It lays a foundation for the automatic control of wiper. The simulation results show that the improved BiSeNet V2 algorithm has an Intersection over Union of 75% compared with the original algorithm. Still, the number of params is reduced by about 8.8%. The number of FLOPS is reduced by about 37.5%, which can improve operational efficiency and facilitate the use of the vehicle-bome system.