A visual detection method based on feature extraction is proposed for three types of powder bed defects in laser selective melting equipment. An adaptive brightness correction algorithm is designed to eliminate the influence of lighting on images caused by uneven brightness from the light source. The defect contours are obtained through image preprocessing, and based on the grayscale information of the images, the defect types are differentiated for feature extraction. Stripe defects are identified using the Hough transform. For the other two types of defects, direction gradient histograms (HOG) features, texture features, and shape features are extracted. The feature vectors are then dimensionally reduced and inputted into the AdaBoost ensemble learning algorithm for training to obtain the classification model. Experimental results show that this algorithm can effectively distinguish the three types of powder bed defects with an identification accuracy of 97.29%. The average detection time is less than 500 ms, enabling fast and accurate recognition.