Feature Fusion-Based Data Augmentation Method for Small Object Detection

In addressing the insufficiencies of feature insertion, inaccurate positioning, and incompatible feature sizes in data augmentation algorithms based on deep learning for detecting microscopic defects on printed circuit boards (PCBs), this paper proposes a novel approach incorporating multiple strategies for small target alignment insertion. First, traditional linear feature extraction methods are transformed into a multiscale comprehensive analysis process. Subsequently, point-to-point matching calculations are converted into region-wise traversals to enhance accuracy and constrain the matching region. Next, geometric correspondences are determined through the computation of a transfer matrix, effectively eliminating perspective distortions. Finally, by constructing a top-down pyramid optical flow module, size limitations are overcome while enhancing features of small target defects. Experimental results demonstrate that this method significantly improves the recognition accuracy of the network model for small target defects on PCB surfaces.