Handwritten text erasion plays a pivotal role in enabling efficient digital archiving and enhancing examination paper reusability within educational systems. Current methods, however, encounter performance bottlenecks due to limited receptive fields of conventional convolution kernels when processing complex page layouts and overlapping text scenarios. To address this challenge, we propose ExamCleaner, a novel handwritten text removal method that integrats the Inception Large Kernel Convolution (ILKC) module and the Context Anchor Attention (CAA) mechanisms. ExamCleaner first expands the model receptive field through the ILKC module and employs parallel multi-scale depthwise convolution to simultaneously capture long-range contextual dependencies and local texture details, thereby effectively distinguishing handwritten and printed text. Then, to alleviate the structural degradation of the original content throughout the erasure phase, the CAA mechanism dynamically anchors key contextual regions (e.g., formula layout, table boundaries) through learnable spatial channel attention to prevent accidental deletion of original content during the restoration process. Experiments conducted on the EnsExam benchmark validate the superior effectiveness of ExamCleaner in handwritten text removal tasks, achieving a remarkable PSNR of 33.45 dB and SSIM of 0.955.