MFFPN: an Anchor-Free Method for Patent Drawing Object Detection

A patent document may contain meaningful drawings that can be used for image retrieval. However, labeling drawing locations manually is time-consuming. Since this work is similar to object detection, some object detection techniques can be employed to facilitate it. In this paper, we propose a new anchor-free object detection method for this purpose. The proposed method contains two parts, namely, max filtering feature pyramid network (MFFPN) and dilated sample selection loss (DSSL). We replace feature pyramid network and path aggregation network by 3D max pooling for multi-scale feature fusion with MFFPN. By using DSSL, we can adaptively select training samples according to the ground truth size. Experimental results show the proposed method can achieve a better performance compared with the state-of-the-art anchor-free methods on Taiwan patent dataset.