Image Content Identification from CNNs with Sparse Sampling

Object detection and classification is a key function for many systems in which input images are examined to find objects and determine their classification and position. Current deep learning networks using CNN perform well when the network is adequately defined in terms of hyper-parameters, and there are sufficient training examples of each of the classification types. We define an approach for extracting sufficient sample data from an image without requiring full frame data set. We demonstrate that sample sparsity with guided object definition can reduce overall sample data requirements while maintaining object classification accuracy.