DenseUNet: Improved image classification method using standard convolution and dense transposed convolution

摘要

U-Net series models have achieved considerable success in various fields such as image segmentation and image classification. However, the decoders in these models often use transposed convolution (TC) from level to level, reducing the representation ability of the features. Therefore, we proposed a DenseUNet model that adopts dense TC to ensure maximum information flow from lower layers to higher ones. DenseUNet comprises four blocks, the output of each block being transposed several times with different strides based on its size. All transposed results of the same size are concatenated together through a skip connection and then fused with the same size result of the first convolutional layer of the corresponding block. Multiscale TC operations restore the feature size at different scales to supplement the important features lost in the pooling layers. By progressively accumulating features from different paths, DenseUNet improves the representational ability of features and enhances the robustness of classification. We evaluated the model on four image classification benchmark datasets – namely, CIFAR-10, CIFAR-100, SVHN, and FMNIST – using three U-Net series models (U-Net, TernausNet, and CrackU-Net) and four classical classification models (VGG16, VGG19, ResNeXt, and DenseNet). The experimental results showed that our model has stable training performance and excellent test accuracy.