<p>Aiming at the problems of unbalanced number of samples and various forms of defect manifestations in wafer defect samples, we propose a wafer defect detection algorithm based on the improved YOLOv9s. Firstly, a channel attention mechanism with inverted parametric structure is designed to make the model efficiently capture global features and fully utilize the key information of wafer defect features. Secondly, in order to reduce the resources required by the model, spd-conv is used instead of the traditional convolutional layer in the downsampling addition module, which reduces the size of the feature map and retains the main feature information. At the same time, dynamic head is used as the detection head of the improved algorithm. Finally, the focal loss function is introduced to adjust the learning process of the model according to the importance of different defect categories to better adapt to the unbalanced data distribution. Experiments are conducted on the Dataset-Wafer data to validate the proposed method, and the experimental results show that the precision, recall, and average accuracy of the method reach 0.971, 0.951, and 0.979, respectively, and the classification accuracy of each category exceeds 0.963, which suggests that the method is reasonable and effective in dealing with the imbalanced wafer defect dataset.</p>