Surface Tiny Defect Detection Based on Gaussian Distribution Modeling and Adaptive Feature Fusion

Surface defect detection in industrial production aims to identify defects on the surface of products, providing information about the category and location. The detection of tiny defects is one of the key challenges in defect detection. Compared to defects at normal scales, tiny defects have fewer effective features and require higher precision in detection box localization. In this paper, we propose a defect detection network based on Gaussian distribution modeling and adaptive feature fusion (GA-NET) to detect tiny defects. Firstly, we introduce an adaptive feature fusion module, allowing the model to automatically learn the importance of different channels. Secondly, we model both the detected rectangular boxes and ground-truth boxes using Gaussian distribution, and calculate the distance between different distributions using the Normalized Wasserstein Distance (NWD), which directs the network's attention toward tiny defects. Finally, we optimize the ℓ1 loss function with the Wasserstein loss to learn better Gaussian distribution. Experimental results on the NEU-DET defect detection dataset demonstrate that our model achieves a mAP of 78.8%, which is a 3.6% improvement over RetinaNet. In comparison to other well-known defect detection algorithms, our model achieves superior detection performance.