Aiming at the problems of fine defects, irregular defects and indistinct defects in the surface inspection of transistors in industry, a transistor defect detection algorithm based on YOLOv11-MRW is proposed. Firstly, this paper fuses the large convolution kernel network UniRepLKNet with the lightweight gated depthwise convolution model Mambaout to replace the C3K2 in the original YOLOv11 model, which enables lightweight and efficient inference for image classification, improves the feature extraction ability at different scales, and increases the receptive field. Two variants of the RFD downsampling module are introduced to enhance the feature extraction ability at different stages, effectively solving the problem of small target information loss caused by traditional convolution. By adopting the WIoUv2 bounding box loss function, the contribution of simple samples to the loss value is effectively controlled at a low level, and the training center of the model is shifted from simple scenes to occluded targets, resulting in a more significant improvement in the detection accuracy of occluded targets. Experimental results show that compared with the basic model, the mAP@0.5 and mAP@0.5-0.95 of YOLOv11-MRW have increased by 1.2% and 10.6% respectively. In practical applications, it has certain research value for the surface detection of fine defects and provides strong support for improving the accuracy of transistors and other chips.