DDNet: Lightweight Model for Enhanced Global Feature Representation

Convolutional neural networks (CNNs) have achieved remarkable results in many fields. However, small convolutional kernels have a limited receptive field, making it challenging to capture global information, which can adversely impact the model's performance on complex tasks. To this end, this paper proposes a model called Dilated-DepthwiseNet (DDNet), which can effectively improve its global representation capability. Its end of the stem uses a <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$3 \times 3$</tex> convolution with a stride of 2, followed by two <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$3 \times 3$</tex> convolutions with a stride of 1, to minimize the loss of information during downsampling. The downsampling layer fuses the feature pyramid with the expansion convolution to enhance feature extraction. Experimental results show that DDNet performs well on Mini-ImageNet. Compared with MobileNetV3-S, the accuracy of DDNet-S0 is 1.31% higher with similar latency. DDNet demonstrates excellent performance in terms of parameters, latency, and accuracy. It is expected to be an efficient backbone network architecture that balances performance and computational efficiency.