In order to reduce the failure probability of the storage tank under thermal radiation conditions, this paper aims to study the post-thermal buckling behavior of fixed headspace thin steel storage tanks under thermal radiation. Finite element software ABAQUS was used to simulate the key factors affecting the thermal buckling response of the target tank (flame height, wind speed, and tank diameter). The results show that the deformation of the tank is directly proportional to the flame height, wind speed and tank diameter. The degree of depression deformation of the target tank will increase with the increase in flame height. The existence of wind speed increases the radial displacement amplitude of the tank wall, and the degree of unevenness of the wall intensifies. Compared with D=5, the deformation degree of the tank and the number of post-buckling occurred more frequently and more intensely in D=10. The reasonable layout of the storage tank position can increase the refractory time of the storage tank, and increase the minimum critical buckling temperature to prevent the occurrence of buckling behavior after heat.

— Image restoration in real-world scenarios, especially in professional fields such as footprint analysis that demand strict detail fidelity, poses significant challenges to the design of network architectures. Most existing methods are tailored for general natural images; when professional images with specific structural textures and complex noise are processed, an optimal balance between denoising and detail preservation is often difficult to achieve. To address this challenge, a novel architecture named Dual-Domain Interactive Fusion Network (DDIFNet) is proposed in this paper, which aims to solve image restoration problems in specific domains.

The core of DDIFNet is a novel Dual-Domain Fusion Module (DDFM), which is designed to process features from the spatial domain and the frequency domain (obtained via Fast Fourier Transform) in parallel. A key design of DDFM is an interactive fusion mechanism: features in the spatial domain (e.g., textures, edges) can guide the filtering process in the frequency domain, while features in the frequency domain (e.g., the spectral distribution of noise) can inversely modulate the feature responses in the spatial domain. This bidirectional interaction mechanism allows the network to adaptively integrate cross-domain information based on local image content.

To evaluate the effectiveness of DDIFNet, core experiments were conducted on a challenging self-constructed forensic footprint image dataset. Experimental results demonstrate that DDIFNet outperforms current state-of-the-art methods (including SwinIR) significantly in restoring key forensic features corrupted by severe noise. Furthermore, to verify its generalization capability, the network was tested on the general denoising benchmark SIDD, and competitive results are achieved. These results prove that the proposed architecture not only specializes in solving specific problems but also maintains good generality, which validates the advancement and practical value of its design.