<p>Defect detection of aluminum quality is crucial to guarantee production quality. In this paper, an efficient multi-scale perceptual enhancement network (EMSPENet) for aluminum defect detection is proposed to achieve a balance between performance and efficiency. First, the InResC2f module is designed to enhance the feature extraction capability, and the VoVGSCSP module is introduced to alleviate the feature dispersion. Second, the Contextual Feature Fusion Module (CFFM) is designed to compensate for the deep feature loss and reduce the background interference to enhance the network sensory field. Finally, in light of the fact that the feature mapping in a PAN encompasses data pertaining to two distinct paths, yet the network exhibits varying degrees of sensitivity to the bottom and top feature mappings, a detail-aware aggregation network (Dp-PAN) is put forth as a means of enhancing the feature information and gradient flow information associated with the various paths, thereby facilitating a more optimal fusion of the local and positional data. The experimental results show that the mAP50 and mAP50–90 of the method were 93.8% and 77.2%, respectively, which were 1.7% and 6.1% higher and FLOPs were 17.7% lower than those of YOLOv8s. The method significantly outperformed the SOTA model in all evaluation metrics.</p>