2026-10-12T23:10:19Z https://api.figshare.com/v2/oai

oai:figshare.com:article/33865883 2026-09-16T17:33:43Z category_146 category_12 category_734 category_931 portal_5 item_type_3 month_year_09_2026

<p>Image quality test results.</p> Wanjie Kang (24998627) Youshun Pan (1390173) Space Science Cell Biology Biological Sciences not elsewhere classified Information Systems not elsewhere classified xception convolutional modules two comparative algorithms structural similarity index outperforming comparative methods experimental results demonstrate dense connection mechanism demonstrating strong robustness computational costs increase bit error rate proposed algorithm maintains reduce computational complexity noise standard deviation complex channel interference strong compression conditions lightweight convolutional modules noise interference experiments proposed algorithm achieves interference capability algorithm achieves model complexity channel features noise ratio test set study proposes steganographic image steganographic capacity secret image scale spatial practical solution peak signal method retains inference stage image quality image imperceptibility findings indicate extraction network encoding network embedding capacity detection capabilities decoding network crucial role covert communication copyright protection carrier image also surpassing 90 compression 7 db <div><p>Image steganography plays a crucial role in covert communication and copyright protection, but existing methods still struggle to achieve a balance between embedding capacity, image imperceptibility, anti-detection capabilities, and model complexity. Under complex channel interference such as compression and noise, the accuracy of secret information recovery easily declines, while computational costs increase. Therefore, this study proposes an image steganography method based on an end-to-end embedding-extraction network and lightweight convolutional modules. First, the carrier image and the secret image are input into the encoding network, where multi-scale spatial and channel features are extracted using improved Inception-V3 and Xception convolutional modules. Second, a RepVGG structure reparameterization mechanism is introduced to reduce computational complexity in the inference stage while maintaining feature representation capabilities. Finally, a dense connection mechanism is combined in the decoding network to enhance shallow feature reuse, and a discriminant network applies adversarial constraints to the carrier image and the steganographic image. Experimental results demonstrate that the proposed algorithm achieves a peak signal-to-noise ratio of 40.7 dB, a structural similarity index of 0.969, and a learned perceptual image patch similarity of 0.055 on the test set, all outperforming comparative methods. In terms of steganographic performance, its bit error rate is only 0.008, and the area under the curve is 0.398, also surpassing the other two comparative algorithms. Additionally, under strong compression conditions of 30, the proposed algorithm maintains a recovery accuracy of 0.82 and controls the bit error rate at 0.38, while under 90 compression, it achieves a recovery accuracy of 0.98 with the bit error rate dropping to 0.23. In noise interference experiments, the method retains a recovery accuracy of 0.96 and a bit error rate of 0.22 when the noise standard deviation is 0.05, demonstrating strong robustness. The findings indicate that this algorithm achieves a balance between performance and efficiency in steganographic capacity, image quality, and anti-interference capability, offering a more practical solution for image steganography applications.</p></div> 2026-09-16T17:33:23Z Dataset Dataset 10.1371/journal.pone.0357380.t003 https://figshare.com/articles/dataset/_p_Image_quality_test_results_p_/33865883 CC BY 4.0