Real-time segmentation network for compact camera module assembly adhesives based on improved U-Net

To meet the real-time and accuracy of glue segmentation in compact camera module images, this paper proposes a real-time segmentation network model based on U-Net. To improve the inference speed of the model, the encoder of the network is redesigned into a three-level feature extraction structure, and the upsampling collocation of bilinear interpolation and sub-pixel convolution is used in the decoder. To enhance information fusion in the context of the network, the context guiding block is embedded in the feature extraction branch of the model. On this basis, improvements to convolutional block attention module and embedded in the jump connection to guide upsampling. The experimental results show that the inference speed of the segmentation network in this paper can reach 227.02 frames per second, which is better than real-time segmentation networks, such as ENet and CGNet. The segmentation accuracy of reinforced glue, dust-trapping glue, and escaped air hole glue is closer to high-precision segmentation networks, such as Deeplab V3+ and U-Net. The network model achieves 96.41% Intersection over Union on the segmentation of the reinforcing glue. Exploring the application of this paper’s network to other objects with simple semantic information is an important direction for future research.