Deep Learning Does Not Generalize Well to Recognizing Cats and Dogs in Chinese Paintings

摘要

Although Deep Learning (DL) image analysis has made recent rapid advances, it still has limitations that indicate that its app-roacn differs significantly from human vision, e.g. the requirement for large training sets, and adversarial attacks. Here we show that DL also differs in failing to generalize well to Traditional Chinese Paintings (TCPs). We developed a new DL object detection method A-RPN (Assembled Region Proposal Network), which concatenates low-level visual information, and high-level semantic knowledge to reduce coarseness in region-based object detection. A-RPN significantly outperforms Y0LO2 and Faster R-CNN on natural images (P < 0.02). We applied YOLO2, Faster R-CNN and A-RPN to TCPs with a 12.9%, 13.2% and 13.4% drop in mAP compared to natural images. There was little or no difference in recognizing humans, but a large drop in mAP for cats and dogs (27% & 31%), and very large drop for horses (35.9%). The abstract nature of TCPs may be responsible for DL poor performance.