Automatically surface defect detection plays a crucial role during the industrial production process. Unfortunately, some special defects, such as weak and strip-like defects, are relatively difficult to classify and localize accurately. In this paper, we propose a novel Location-Aware Guidance Network for weak and strip defect detection, termed as LAGNet. To enhance the feature representation of weak defects, we introduce the Location Activation Map (LAM) by visualizing the confidence score map that indicates the probability of the existence of an object in each region. The LAM and RGB images are fed into the network in a parallel manner for feature extraction, and then we fuse these two branches via a Location Guidance Block (LGB) that inherently encodes comprehensive and complementary information for detection. Additionally, a Strip Convolution Enhancement Module (SCEM) is presented using the depthwise strip convolutions with long but narrow kernels and attention mechanism and plugged into the detection neck to model the long-range dependencies along both horizontal and vertical spatial directions, thus improving the detection performance of anisotropic defects with banded structures. Notably, LAGNet achieves the top-ranking results on two popular steel benchmarks and significantly outperforms the baseline network YOLOv5: 85.5% mAP (vs. 76.2%) on NEU-DET and 76.6% mAP (vs. 66.0%) on GC10-DET.