An Wavelet Steered network for efficient infrared small target detection

Abstract

This paper introduces an innovative approach for infrared small target detection (IRSTD). Our model is primarily motivated by the integration of Discrete Wavelet Transform (DWT) into a convolutional neural network (CNN), combining the classical multi-scale analysis method with the deep learning (DL) framework. Firstly, it is argued that detail loss may be induced by downsampling operations such as average or maximize pooling, and the interpolation of upsampling is inefficient in increasing useful information. Therefore, Haar DWT and inverse DWT (IDWT) are embedded for lossless downsampling and upsampling, facilitating more effective feature extraction within the CNN. Secondly, a hybrid attention mechanism, referred to as Wavelet Steered Transformer (WST), is designed to fully enhance the DWT features both spatially and across channels. This mechanism consists of two key improvements: (1) Channel-wise Transformer: We propose adapting a channel-wise Transformer to enhance semantic information and suppress background clutter and noise. This enhancement ensures that the targets and background features are distributed across different channels, thereby boosting detection performance. (2) Dilated-Gate Convolutional Module: A dilated-gate convolutional module is employed to enhance spatial location accuracy. Unlike previous methods for location extraction, this module uses a combination of different kernel sizes and dilation rates to improve spatial accuracy. Experimental results on benchmark datasets showcase the supervisor performance of the proposed method. The code and data for this paper will be released at https://github.com/Fortuneteller6/WaveTD once the paper is accepted.