Abstract: China is a large country in the manufacture of aluminum profiles, and the quality inspection of aluminum profiles is of great significance. Aiming at the problems of low detection efficiency and relatively weak stability of traditional manual visual inspection methods, the feature extraction of the single YOLOv3 method is not prominent, and the detection accuracy is limited, this paper proposes a method for detecting surface defects of aluminum profiles based on image fusion and YOLOv3. First, image enhancement and spatial filtering methods are used to preprocess the original image to obtain a processed image; then the feature extraction and matching ideas in SLAM are used for reference to extract and match the original image and the processed image; then image fusion is performed to obtain the final processed image; subsequently, the K-means algorithm is used for clustering and tuning parameters optimization. Finally, a single-stage object detection model YOLOv3 is used to detect the surface defects of aluminum profiles. An end-to-end full convolutional neural network is used to complete the input from the original image to the output of the object categories and confidence in the Bounding box and box. The experimental results show that the average success rate of surface defect classification detection by this method is 98.33%, which is 3.75 percent points higher than the single YOLOv3 method; the mAP value of the validation set is 88.81%, which has an increase of 4.18 percent points. It has stronger feature extraction and generalization capabilities, and can accurately detect surface defects, perform classification and localization.

Key words: image fusion, deep learning, SLAM, feature extraction and matching, damage detection, YOLOv3； K-means algorithm