The study of the similarity evaluation and retrieval of patent documents is critical not only for the efficient management of patent literature, but also for the rapid and effective collection of information in industrial and technological fields. Patent drawings visually represent the outcomes of technological advancements and innovations, but have not been given as much importance as texts in the past.

This study evaluated the similarity of patent drawings for effective retrieval using the representative deep-learning model, ResNet-50, and the traditional computer vision algorithm, scale-invariant feature transform (SIFT). First, a classification experiment using 10,827 patent drawings was conducted to evaluate the similarity of the visual types, achieving a classification performance with an accuracy exceeding 95%. Second, a retrieval experiment using 5,000 technical drawings was conducted to compare the features of ResNet and SIFT based on their similarity. Finally, the retrieval and matching performances of ResNet and SIFT were evaluated using 50 original data samples and 4,800 augmented data samples created by various forms of editing. ResNet demonstrated an average matching performance of 72.54%, whereas SIFT achieved an average matching performance of 86.71%.

The findings reveal that, unlike ResNet-50, which compares similarity using the entire image information, SIFT evaluates similarity based on attribute information, such as key points within the image. Consequently, ResNet is advantageous for identifying visually similar images, whereas SIFT excels in identifying identical images.