Delving deep into light-weight salient object detection
Jia-wen Xiao, Jiekang Feng · 2021
Salient object detection is still a hot topic at the moment. Based on the pixel-by-pixel prediction of the image, the salient object detection model often has a large number of parameters, which brings problems such as high latency, large-scale computing and storage difficulties. In this article, we try to explore the application of the idea of structural reparameterization in real-time salient object detection, trying to alleviate the contradiction between the above three problems and accuracy and focusing on the important feature of optimizing the reasoning speed of the network. Based on the above ideas, we propose a plug-and-play flexible convolution with a new structure. The convolution module can be converted to a normal 3x3 convolution layer in the inference stage, so the inference speed will be greatly improved. Using the plug-and-play convolution module and pooling together, we construct a model with high inference speed.