US 11,727,056 B2
Object detection based on shallow neural network that processes input images
Igal Raichelgauz, Tel Aviv (IL); and Roi Saida, Acco (IL)
Assigned to Cortica, Ltd., Tel Aviv (IL)
Filed by Cortica, Ltd., Tel Aviv (IL)
Filed on Mar. 31, 2020, as Appl. No. 16/835,333.
Claims priority of provisional application 62/827,122, filed on Mar. 31, 2019.
Prior Publication US 2020/0364474 A1, Nov. 19, 2020
Int. Cl. G06V 20/58 (2022.01); G06F 16/901 (2019.01); G06N 3/08 (2023.01); G06N 3/04 (2023.01); G06T 11/20 (2006.01); G06V 40/10 (2022.01); G06F 18/2431 (2023.01)
CPC G06F 16/901 (2019.01) [G06F 16/9014 (2019.01); G06F 18/2431 (2023.01); G06N 3/04 (2013.01); G06N 3/08 (2013.01); G06T 11/20 (2013.01); G06V 20/584 (2022.01); G06V 40/10 (2022.01); G06T 2210/12 (2013.01)] 30 Claims
OG exemplary drawing
 
1. A method for object detection, the method comprises:
receiving an input image by an input of an object detector; and
detecting, by the object detector, objects that appear in the input image; wherein the detecting comprises:
processing the input image by a shallow neural network to provide a shallow neural network output; and
determining, by one or more region units, bounding boxes information related to bounding boxes that surround at least some of the objects that appear in the input image, based on the shallow neural network output;
wherein for each one of the bounding boxes, the bounding box information comprises coordinates of the bounding box, objectiveness of the bounding box and a class of an object that is surrounded by the bounding box; wherein the coordinates indicate a location, a height and a width of the bounding box; and wherein the objectiveness provides a confidence level that the object exist;
wherein the shallow neural network output comprises multiple convolutional layers and multiple pooling layers; wherein the multiple convolutional layers comprise convolutional filters having a kernel that has more than nine elements;
wherein an output of a certain convolutional layer that is located at a beginning of the shallow neural network has significantly fewer elements than an output of the first convolutional layer of the shallow neural network.