In order to improve the accuracy of complex compound analysis in comprehensive two-dimensional gas chromatography (GC×GC) data, an improved model named AP-U-net was proposed based on the U-net semantic segmentation framework to investigate the precise segmentation of streak noise. Streak noise was characterized by low contrast and multi-scale variations, which makes it challenging to segment this target in complex backgrounds. To address the issues of low feature propagation efficiency, information loss caused by pooling operations, and difficulty in capturing global context in the original U-net, residual connections were introduced to enhance gradient flow, Haar wavelet decomposition (HWD) was used to replace downsampling, and a selective kernel (SK) attention mechanism was incorporated to improve global information modeling. The results show that the improved AP-U-net model achieved a mean intersection over union (MIoU) of 81.44%, accuracy (Acc) of 98.69%, class pixel accuracy (Cpa) of 83.22%, and a Dice coefficient of 78.21% on the GC×GC dataset. Compared with the original U-net model, the MIoU increases by 3.83%, and the Dice coefficient increases by 5.7%. It is concluded that the AP-U-net model effectively segments streak noise and shows certain advantages over some mainstream segmentation models, indicating good application potential.