Method of board lumber grading using deep learning techniques

Method of detecting characteristics of wood material includes creating a model definition of multiple classes of wood quality of multiple species where the wood quality characteristics are learned from images where specimens each having major surface areas and the images are acquired from multiple channel sensors having outputs produced by automatic scanning of the specimens and identifying the characteristics. The specimen images are represented by layers of input layer pixel data derived from raw image data produced by automatic scanning of the multiple wood specimens channel sensor outputs and each layer represents a different one of the channel sensor outputs. The input layer pixel data correspond to sets of pixels representing regions of each of the multiple specimen images, the regions encompassing smaller surface areas than major surface areas. The model further includes using the input layer pixel data representing the multiple classes of the wood quality characteristics within specified boundaries of the regions at identified locations. The method receives the input layer pixel data and a machine learning framework supports a training processing unit in which is performed a set of deep learning algorithms developed to train a machine learning-based convolutional neural network on semantic segmentation, the algorithms performing semantic segmentation on the input layer pixel data to determine edges in and network learned weights for collections of pixels in the sets of pixels, the collections of pixels encompassed by the edges and corresponding to the regions of each of the multiple wood specimens. The milled board image data includes input layer pixel data and which is then applied to the convolutional neural network operating on an inference processing unit to form a series of probability maps , each probability map in the series corresponding to a different one of the multiple classes of wood quality characteristics so that each milled board pixel of the milled board pixels has a probability value for each of the multiple classes of wood quality characteristics whereby the inference processing unit derives a solution identifying which ones of the milled board pixels belong to one or more of the multiple classes of the wood quality characteristics and specifying the classes of wood quality characteristics to which the identified milled board pixels belong.