Prompt Description

The input image is first denoised by a denoising module, and then enters the PSF feature frequency domain fusion module to realize deconvolution. The latter has two deconvolution branches, one for image-level deconvolution and one for feature-level deconvolution. The use of dual branches takes into account two core problems in complex marine underwater imaging scenarios. On the one hand, image-level deconvolution is susceptible to noise amplification and artifact interference. Deconvolution is essentially a high-frequency enhancement process. When the signal-to-noise ratio (SNR) is low (such as in deep-sea low-light, high-scattering environments), and when there are errors in the PSF, sensor noise and quantization errors will be significantly amplified, resulting in ringing effects, overshoot, or false textures in the reconstructed image. These artifacts may be misjudged as target features by subsequent recognition networks, which may reduce recognition robustness. On the other hand, feature-level degradation is not explicitly modeled. Even if the image looks "clear", the high-level semantic features extracted by the deep neural network may still be distorted due to the original degradation (for example, edge response offset, texture energy attenuation). If only image-level restoration is relied upon, it cannot be guaranteed that the discrimination in the feature space will be restored. For the image-level deconvolution branch, the image is deconvolved to generate the first initial reconstructed image; for the feature-level deconvolution branch, the image is passed through a feature extraction module (FM) to obtain a series of feature maps, which are Wiener inverse filtered and restored to the second initial reconstructed image through a feature reconstruction module. After that, the two initial reconstructed images are superimposed and input into a multi-scale residual fusion module, and the final clear reconstructed image is output. Please help me draw the overall network structure diagram based on this.