Abstract: To address the issues of high complexity and limited effective receptive field in current infrared target detector backbone network algorithms, this paper proposes a hybrid detection algorithm, WTMLLAYOLO, based on wavelet transform convolution and linear attention. First, a multi-scale wavelet feature extraction module is constructed to expand the receptive field and enhance low-frequency structural information response without significantly increasing model complexity, thereby improving the representation ability of contour and shape features. Second, a non-recursive bi-branch linear attention mechanism is designed to model the global context with parallelizable linear complexity and promote local-global information interaction, reducing the risk of missed detections of occluded and near-range targets, and further improving detection accuracy. Finally, the cross-layer fused features are recalibrated by decoupling the detection head to eliminate semantic conflicts between multi-scale features. Experiments on the FLIR and MSRS datasets show that the algorithm achieves a 3.5% improvement in accuracy compared to the benchmark model while reducing computational cost by 0.3G (FLOPs), significantly improving detection accuracy and robustness.