Dense Nested Network Based on Position-Aware Dynamic Parameter Convolution Kernel for Infrared Small Target Detection

Abstract

Infrared small target detection is widely used in military and civil security fields. Most of the existing infrared small target detection algorithms focus on improving detection accuracy. However, the lightweight and generalization capabilities of the network have not improved significantly. To improve the network's generalization ability and lightweight while main- taining accuracy, a dense nested network based on position aware dynamic parameter convolution kernel (Par-DPC DNNet) was proposed. Firstly, a generation module for the dynamic parameter convolution kernel was proposed. The convolution parameters are generated dynamically based on the instance. This can realize the network's local dynamic change for unfixed small targets and effectively improve the network's generalization ability. Moreover, this can achieve good detection performance on the basis of network depth reduction. Secondly, in order to apply the dynamic parameter convolution kernel in the network on a large scale, inspired by the patch fragment mode in the VIT structure, a new downsampling method was proposed. The dynamic multiples of downsampling are realized by the feature map fragment combined with the convolution method without information loss. Finally, the dynamic parameter convolution kernel and new downsampling were integrated into the dense nested network, and a new nonlinear feature extraction method was adopted. It improves network accuracy while maintaining network generalization and structure optimization. The effective- ness of the proposed method was verified using two data sets. The experimental results demonstrate that the detection performance of Par-DPC DNNet is superior to the existing state-of-the-art methods even if the number of internal layers of resblock is reduced to two.