基于自适应深度特征编码器的手机屏幕缺陷检测方法，首发论文 -凯发网

a mobile screen defect detection method with adaptive deep feature encoder

• 1、school of automation, guangdong university of technology, guangzhou 510006, china

abstract：in this paper, we propose a novel mobile screen defect detection framework based on the u-net network. the u-net network has been widely utilized in the field of defect detection due to its simple architecture and high accuracy. however, we observed that the traditional u-net encoder structure is excessively simplistic, consisting of only a few convolutional layers, which hinders the network from adequately capturing crucial features of the target. additionally, it lacks a rational mechanism to leverage deep and shallow-level feature information, resulting in a lack of accuracy in the segmentation results, particularly in the detailed regions. moreover, the u-shaped structure's bottom feature map has a large number of channels, and the direct connection between the encoder and decoder increases the computational complexity, making it unsuitable for real-time requirements. to address these issues, we introduce a dynamic stacked encoder to replace the traditional u-net encoder. this dynamic stacked encoder can dynamically adjust the utilization of deep and shallow-level feature maps based on specific targets, thereby achieving more accurate segmentation performance. furthermore, to simplify the network structure and reduce computational complexity, we incorporate a path fusion module. this module fuses the encoding-decoding structure located at the bottom of the u-net into a single path, significantly reducing the number of parameters and inference time. finally, we introduce a boundary loss function in the loss function and adopt a weighted joint training strategy to balance the network's attention towards defect boundaries and main regions, thereby comprehensively enhancing the segmentation accuracy of the network.

keywords：