Knowledge Guided Deep Learning for General-Purpose Computer Vision Applications

This research targets general-purpose smart computer vision that eliminates reliance on domain-specific knowledge to reach adaptable generic models for flexible applications. It proposes a novel approach in which several deep learning models are trained for each image. Statistical information of each trained image is then calculated and stored with the loss values of each model used in the training phase. The stored information is finally used to select the appropriate model for each new image data in the testing phase. To efficiently select the appropriate model, a kNN (k Nearest Neighbors) strategy is used to select the best model in the testing phase. The developed framework called KGDL (Knowledge Guided Deep Learning) was evaluated and tested using two computer vision benchmarks, 1) ImageNet for image classification, and 2) COCO for object detection. The results reveal the effectiveness of KGDL in terms of accuracy and competitiveness of inference runtime. In particular, it achieved 94% of classification rate in ImageNet, and 92% of intersection over union in COCO dataset.