上海科技大学知识管理系统

本发明提出一种嵌入式智能加速器的图像反射消除方法和装置，包括：获 取已标记图像标签的反射图像作为训练数据；构建基于 Swin Transformer 的网络架构的反射消除模型，该网络架构包含用于捕捉图像中的长距离依赖关系 的自注意力机制；在嵌入式智能加速器所属的智能 AI 芯片平台搭建该反射消 除模型；该嵌入式智能加速器通过该反射消除模型对该训练数据进行反射消除， 得到反射消除图像，根据该反射消除图像和该图像标签构建损失函数，训练该 反射消除模型；将待消除反射的图像输入嵌入式智能加速器，通过训练完成后的反射消除模型对待消除反射的图像进行反射消除处理，得到反射消除图像作 为图像反射消除结果。

The present invention proposes an image reflection cancellation method and apparatus for an embedded intelligent gas pedal, comprising: obtaining a reflection image labeled with an image tag as training data; constructing a reflection cancellation model based on the network architecture of Swin Transformer, which includes a self-attention mechanism for capturing long-distance dependencies in the image; constructing the reflection cancellation model on the intelligent AI chip platform belonging to the embedded intelligent gas pedal; the embedded intelligent gas pedal cancels the reflection of the training data through the reflection cancellation model to obtain a reflection cancellation image; constructing a loss function based on the reflection cancellation image and the image tag; and training the reflection cancellation model. The reflection cancellation model is constructed on the intelligent AI chip platform belonging to the embedded intelligent gas pedal; the embedded intelligent gas pedal removes reflections from the training data through the reflection cancellation model to obtain a reflection cancellation image; the reflection cancellation model is trained by constructing a loss function based on the reflection cancellation image and the image labels; the image to be reflected is input to the embedded intelligent gas pedal; the image to be reflected is processed through the trained reflection cancellation model to obtain a reflection cancellation image Input the image to be removed into the embedded intelligent accelerator, process the reflection removal by the trained reflection removal model, and obtain the reflection removal image as the result of image reflection removal.