Images, Pre-trained Networks, and Information

This research explores a workflow for describing the relationship between machine learned task complexity and image quality using an image quality metric. The outcome provides strong evidence for the feasibility of an automated analytic scale. To perform these critical experiments, we leverage the current state-of-the-art machine learning detection and classification algorithms that utilize convolutional neural networks (CNNs). The CNN generation of machine learning algorithms has two attractive attributes over previous generations that we leverage in this exploration: 1. Ingest raw imagery rather than hand crafted problem specific features; and 2. Contain a huge volume of computational neurons that produce fine grained discriminations. Previous studies relating task complexity to image quality focused on human perception. These studies do not effectively transfer to automation because 1. human perception can address any task solved at coarser image quality can also be solved with higher image quality and 2. human perception measures do not incorporate false-positive rates into their scale.For this research, we generate image features from pre-trained networks across different image classification tasks with increasing complexity. Those image features are images inferred through a VGG16 CNN model trained with ImageNet weights. We analyze CosmiQ Rare Planes dataset with 2, 4, 8, and 16 classes. Then, we apply the Euclidean distance metric across class means to characterize task complexity. Though we expect better metrics to exist, the workflow shows that an automated analytic scale is feasible for image detection and discrimination tasks.