High-density polyethylene (HDPE) is a novel thermoplastic material for fishing vessel construction, whose welding quality is one of the primary factors ensuring the safety of HDPE fishing vessels. To address the challenges in HDPE fishing vessel welds defect detection, including high similarity between defects and background as well as weak small-target features, this study proposes an improved ACA-YOLOv8(Adown-CCFM-AC-mix-YOLOv8) object detection algorithm.The proposed method employs an Adaptive Downsampling(ADown) strategy to effectively preserve defect features, enhances multi-scale feature representation through a Cross-scale Consistent Feature Fusion Network(CCFM), and incorporates a Self-attention and Convolution Mixed(AC-mix) mechanism during feature fusion to improve small target detection capability.Experimental results demonstrate that the improved model maintains lightweight characteristics while achieving an average detection accuracy of 98.9%, representing 3.2%  improvement over the baseline model. Additionally, it reduces parameters by 43.5% and computational load by 2.0G. This algorithm better meets the computational requirements for HDPE fishing vessel welds defect detection in industrial production environments.