Burrs are thorny protrusions characterized by uneven edges or excess debris on the metal surface. In the aluminum profile extrusion industry, while a few burrs do not significantly affect the product’s functionality, they can adversely impact the measurement of critical parameters such as inner and outer diameters. Traditionally, burrs on aluminum profiles have been manually removed before measurement. However, it is feasible and highly efficient to directly remove burrs from cross-sectional images of aluminum profiles with burrs and measure certain parameters using computer vision methods, considering the labor-intensive nature of manual removal and the sporadic occurrence of burrs. Therefore, designing an automatic and robust method to deburr cross-sectional images of aluminum profiles with burrs is imperative. This paper investigates the efficacy of multiple image restoration models for deburring cross-sectional images of aluminum profiles with burrs. Acquiring numerous pairs of cross-sectional images of aluminum profile from the real world is prohibitively expensive. Therefore, we propose a semi-automatic method to synthesize cross-sectional burrs and non-burrs images for training deburring models. Additionally, we design an explainable mechanism called Axial Counting Attention to enhance burr removal effectiveness. This mechanism is effective and capable to focus on the edge areas with burrs in convolution-based networks. Experimental results demonstrate that our simple deburring model effectively removes burrs comparable to manual methods or even surpasses human performance, while maintaining competitive computational costs and inference times.