Injection-molded products may to have a variety of defects in production. Failing to detect and fix the defects may reduce product quality and lead to safety issues. An injection-molded product defect detection model, IMP-DETR, is proposed to address the challenges of diversity, small size, and complex background in injection-molded products. The model constructs a feature extraction backbone network with the iRMB module to extract key information and reduce interference from irrelevant backgrounds while maintaining lightweight. The SOFP feature fusion network is used to capture rich texture information from small objects to improve the detection performance of fuzzy and small-sized defects. Additionally, the Conv3XC-Fusion module is designed to resolve the problem of integrating multi-scale features, improving the stability of detection. Due to the lack of publicly available datasets for injection-molded product defects, we constructed a custom dataset containing 2500 defect images. The experimental results indicate that the mAP of the IMP-DETR model reaches 82.4%. Compared to other benchmark object detection models, IMP-DETR demonstrates superior detection performance and a smaller model size, which is suitable for application in real scenarios.