Abstract: ‍ ‍Real image denoising aims to remove noise distortion from noisy digital photograph images to improve their visual quality. The state-of-the-art real image denoising models generally rely on sample set of pixel-wise corresponding clean-noisy image pairs， which are difficult to collect. To address this challenge， we propose a content- and quality-constrained real image denoising network （CQDeNet）， a network for real image denoising constrained by content and quality， trained solely on image-level labeled samples to achieve effective denoising. CQDeNet consists of two sequential sub-networks： the image content-constrained network and the image quality-constrained network. The content-constrained network is trained in a self-supervised manner using only noisy images to restore the content information by extracting the latent clean image signal from the noisy images. The quality-constrained network utilizes a no-reference image quality assessment to constrain the quality of the generated images， collaboratively guiding the optimization direction of the content-constrained network. By combining these two sub-networks， CQDeNet overcomes the limitation of the zero-mean assumption of noise residuals in current unsupervised methods， enabling the model to have stronger generalization ability and scalability. Test results show that the proposed method could effectively remove the noise from real images. The average PSNR ​​values obtained on the SSID and DND datasets are 34.83 dB and 37.21 dB， respectively.