Ultrasonic inspection of coarse-grained steels is a common challenge in various industrial fields. This task is often difficult because of acoustic scattering that creates structural noise in the ultrasonic signals and images. This drives inspections using low-frequency probes at the cost of a lower resolution of standard delay and sum (DAS) imaging techniques, such as the well-known total focusing method (TFM). The purpose of this paper is to present and evaluate the performances of an image reconstruction technique that aims at improving the resolution when inspecting industrial coarse-grained materials. An image deconvolution problem (with spatially varying blur) is formulated, relying on a forward model that links the TFM image to the acoustic reflectivity map. A particular attention is paid to the estimation of the PSF used for the deconvolution approach in an experimental context. The experiments are based on an austenitic-ferritic sample insonified using array probes at 3 MHz and 5 MHz placed in contact. The goal is to resolve two close reflectors corresponding to side drilled holes (SDH) with diameter 0.4 mm spaced by 0.4 mm edge to edge and positioned at different depths (10, 20, 30, 40 mm). This configuration corresponds to a critical case where the distance between the two reflectors is significantly inferior to the Rayleigh distance, that is the resolution limit of a DAS imaging system. These are typical cases where the employed frequency is actually too low and where a higher frequency probe should be used, which is not possible in practice, because it would affect the detection capability due to higher noise level. As predicted by the Rayleigh criterion, TFM is not able to separate the reflectors. The proposed image reconstruction method successfully resolves the majority of the reflectors with a rather accurate distance estimation. In the context of coarse-grained structure inspection, this approach enables the use of low-frequency probes, in order to improve the signal-to-noise ratio, while keeping high resolution capability.