| CPC G06T 7/136 (2017.01) [G06T 7/0012 (2013.01); G06T 7/11 (2017.01); G06T 7/174 (2017.01); G06V 10/26 (2022.01); G06V 10/764 (2022.01); G06V 10/7715 (2022.01); G06V 10/776 (2022.01); G06V 10/82 (2022.01); G06V 10/95 (2022.01); G06V 20/68 (2022.01); G06T 2207/20081 (2013.01); G06T 2207/20084 (2013.01); G06T 2207/30188 (2013.01); G06T 2207/30242 (2013.01)] | 19 Claims |
|
1. A processor-implemented method comprising the steps of:
receiving, by one or more processors, an image from a temporal set of images of at least a portion of one or more crops being monitored for a pre-defined event;
detecting one or more Regions Of Interest (ROIs), by the one or more processors, as localized bounding boxes in the received image based on the pre-defined event using one or more context sensitive pre-trained models associated with the pre-defined event for the one or more crops being monitored, wherein the one or more ROIs correspond to a detected event;
performing unsupervised segmentation of the one or more ROIs, by the one or more processors, using a Convolutional Neural Network (CNN) to obtain a segmentation map having predicted labels based on spatial continuity of pixels comprised within each of the ROIs;
obtaining superpixels from the received image, by the one or more processors, using a superpixel generating algorithm; and
iteratively auto-correcting, by the one or more processors, the obtained segmentation map by:
updating labels of the pixels in the one or more ROIs using boundaries of the obtained superpixels by using unsupervised segmentation and computing features on one or more regions from the received image to obtain disjoint partitions of the received image;
computing a cross entropy loss between the predicted labels by a last convolution layer of the CNN and the updated labels;
back propagating the computed cross entropy loss to the CNN,
until a change in the cross entropy loss from a previous iteration is less than a predetermined threshold;
updating weights of the CNN based on the updated labels in each iteration of the auto-correction of the obtained segmentation map;
identifying the CNN at the end of the iterative auto-correction of the obtained segmentation map as a new pre-trained model; and
updating a model library comprising the one or more pre-trained models with the new pre-trained model, wherein each of the one or more pre-trained models is associated with a corresponding architecture definition, and
inferencing one or more images by the one or more pre-trained models on an edge component where computation resources are limited and deploying the one or more pre-trained models in a model library for making offline image inferences, and the offline image inferences are sent with one or more received images to the model library on availability of a connection.
|