Indoor Scene Recognition with Convolution Supervision and Fusion Deep Network
Ruixin Wang, Lisha Xiao, Yan Qin, Xin Wang · 2019
In order to increase the recognition accuracy of indoor scenes, a network called Convolution Supervision and Fusion Deep Network (CSFDN) based on AlexNet is introduced. Convolutional layers with large convolution kernels are replaced by cascaded ones with smaller kernels and stride. Features of a middle convolutional layer are integrated into that of the high layer to add local information through a fusion branch. A supervision branch is added to the middle convolutional layer to shorten the distance of back propagation and prevent gradient disappearance when training the network. The experimental results on three different datasets indicate that the CSFDN has a higher recognition accuracy than the AlexNet and good generality in indoor scene recognition, which proves the effectiveness of the improved architecture designs.