摘要
Infrared thermography is an efficient non-destructive testing technique. However, thermographic images acquired in practical inspection scenarios are often affected by environmental disturbances, leading to enhanced noise and non-uniform background responses. To address these challenges, this study proposes, for the first time, a defect-aware channel-selection convolutional autoencoder (DC2AE) to improve the visual detection of internal defects in composite materials. The proposed method introduces an adaptive weighting strategy into the multichannel feature extraction process of a convolutional autoencoder, enabling the network to differentially model defect-related responses embedded in different channels. In addition, a defect-aware channel selection mechanism is developed to automatically identify the thermographic response image with the highest defect sensitivity from multi-channel latent feature representations. Furthermore, a comparative skewness-based image selection strategy is used to quantitatively evaluate the thermal response distributions of different candidate channels. The proposed method is also compared with principal component thermography, demonstrating its superiority in defect enhancement and background suppression. Visual analysis of randomly selected channel images from a single output further shows that the channel attention mechanism effectively guides the network to focus on defect- aware regions and generates thermographic features with differentiated representation capabilities across channels. Finally, experiments on three carbon fiber reinforced polymer specimens with representative defects demonstrate that the proposed method produces clearer defect contours and better image-based detection performance than the conventional convolutional autoencoder.