Convolutional networks can be improved by adding direct connections between layers, leading to better accuracy and efficiency. The Dense Convolutional Network (DenseNet) connects each layer to every other layer, allowing for better information flow and reuse of features. This approach helps to solve the vanishing-gradient problem and reduces the number of parameters needed. DenseNets outperform other models on object recognition tasks like CIFAR-10, CIFAR-100, SVHN, and ImageNet, while using less memory and computation. The code and pre-trained models are available for use.