Open-World Object Detection （OWOD） extends the object detection task to real and variable environments， and requires models to identify known and unknown objects accurately and learn new knowledge gradually. In response to the low recall for unknown classes and the problem of false identification in the existing OWOD methods， a Gradient-Discriminative and Feature Norm-driven OWOD （GDFN-OWOD） network model was proposed. To address the issue of low recall for unknown classes， a Gradient-Discriminative Representation Module （GDRM） was proposed， which uses the gradient difference from backpropagation to distinguish unknown classes from the background accurately， thereby improving the recall for unknown classes. In addition， a Graph Segmentation-based Bounding box Clustering （GSBC） algorithm was introduced to model the determination of object bounding boxes as a graph decomposition problem， thereby reducing redundant bounding boxes， and thus reducing the computational complexity of the model. To tackle the problem of false identification for unknown classes， a FeatureNorm-Based Classifier （FN-BC） was employed to select the best-performing convolutional layer to identity known and unknown classes for higher identification precision. Experimental results on M-OWODB dataset show that compared with the best performance of comparison models in tasks T₁， T₂， and T₃， GDFN-OWOD has the recall for unknown classes increased by 1.1， 2.1， and 0.9 percentage points， respectively， and the Absolute Open-Set Error （A-OSE） reduced by 35.1%， 28.7%， and 12.2%， respectively. It can be seen that compared with the existing OWOD methods， the proposed method alleviates the problems of low recall for unknown classes and false identification effectively.