System and method for classifier training and retrieval from classifier database for large scale product identification

Abstract

The disclosure relates to systems and methods for real-time detection of a very large number of items in a given constrained volume. Specifically, the disclosure relates to systems and methods for retrieving an optimized set of classifiers from a self-updating classifiers' database, configured to selectively and specifically identify products inserted into a cart in real time, from a database comprising a large number of stock-keeping items, whereby the inserted items' captured images serve simultaneously as training dataset, validation dataset and test dataset for the recognition/identification/re-identification of the product.

What is claimed: 
     
         1 . A system for automated product identification in a shopping cart, comprising:
 a. a cart having a front wall, a rear wall and two side walls forming an apically open container with a base;   b. a plurality of imaging modules coupled to the cart, adapted to, at least one of image an item inserted into the cart, and image an area of interest outside the cart;   c. a central processing module in communication with the plurality of imaging modules, the central processing module comprising a processor and being in further communication with a non-volatile memory having thereon:
 i. a classifiers' database; 
 ii. a product characteristics database; and 
 iii. a processor readable media comprising a set of executable instruction, configured, when executed to cause the processor to retrieve from the classifier database a set of a plurality of classifiers, 
   wherein each set of classifiers is configured to identify a single product inserted into the from an entire warehouse, storeroom, shop, grocery store, supermarket, or a combination comprising one or more of the foregoing.   
     
     
         2 . The system of  claim 1 , wherein the set of executable instructions is configured, when executed, to cause the processor further to:
 a. using the imaging module, acquire an image of each inserted product;   b. extract a predetermined number of features from the image; and   c. identify the features extracted.   
     
     
         3 . The system of  claim 2 , wherein the features extracted by the imaging module from the inserted product is at least one of: color pallet designator, a shape designator, volumetric size, weight, a key-word, a graphic element, a scale and rotation invariant feature, and a deep feature synthesized from at least one of: a predetermined CNN, and a plurality of the product characteristics. 
     
     
         4 . The system of  claim 3 , wherein the color pallet designator is HSV, CIE-LAB, CIE-XYZ, CIE-LCH or a color pallet designator comprising a combination of the foregoing. 
     
     
         5 . The system of  claim 4 , wherein the set of executable instructions, prior to extracting the predetermined number of features, are configured to preprocess the image of each product inserted to the cart by reducing the number of colors of the product by quantizing colors using at least one of: K-means, median-cut, octree, variance-based, binary splitting, greedy orthogonal bipartitioning, optimal principal multilevel quantizer, minmax, fuzzy c-means and their combination. 
     
     
         6 . The system of  claim 5 , wherein the set of executable instructions, when executed, are configured to cause the processor to:
 a. produce a color histogram of the image color palette; and   b. identify the color characteristic of the inserted product, through retrieving relevant classifiers by associating the color histogram of the inserted product to the classifier's-associated color-histograms by determining similarity that is based on cross histogram bin distance analysis.   
     
     
         7 . The system of  claim 1 , wherein the imaging module comprises a RGB-D camera, and wherein the set of executable instructions, when executed, are further configured to cause the processor to:
 a. using the RGB-D, determine the size of the product;   b. using a load cell module included with the system, determine the weight of the product: and   c. combine the size and the weight to a single parameter.   
     
     
         11 . The system of  claim 1 , wherein the set of executable instructions, are further configured, when executed, to cause the processor to identify and retrieve at least one of: a logo, a watermark, a key word, a graphic symbol, and their combination using at least one of: a single-shot MultiBox detector (SSD) neural network, a Regional convolutional neural network. (RCNN), a Fast-RCNN, a Faster-RCNN, and a You Only Look Once (YOLO) neural network. 
     
     
         12 . The system of  claim 1 , wherein the set of plurality of the classifiers identifying a single product are selected by applying a nearest-neighbors algorithm and approximated nearest neighbors (ANN) algorithm to the classifiers associated with the product independently; and selecting the classifiers selected by both algorithms. 
     
     
         13 . The system of  claim 2 , wherein product shapes are encoded numerically. 
     
     
         14 . The system of  claim 1 , wherein the classifiers' database is formed by training the plurality of classifiers on products sharing at least one characteristic, wherein formed classifiers are associated with these set of characteristics for later retrieval. 
     
     
         15 . The system of  claim 14 , wherein the set of executable instructions, are configured, when executed, to cause the processor to associate each of the product characteristic to a classifier or a set of classifiers. 
     
     
         16 . The system of  claim 14 , wherein the set of executable instructions, are configured, when executed, to cause the processor to retrieve a classifier or a plurality of classifiers from the classifier database, by using each of the product characteristics or their combination as a key for retrieval. 
     
     
         17 . The system of  claim 6 , wherein the cross-histogram bin distance is measured using at least one of: Kullback-Leibler (KL)-Divergence, Bhattacharyya distance, and Chi-square (X 2 ) distance. 
     
     
         18 . The system of  claim 15 , wherein the set of executable instructions, are configured, when executed, to cause the processor to train products' recognition and/or identification algorithm, by: grouping products sharing at least one characteristics. 
     
     
         19 . The system of  claim 1 , wherein the non-volatile memory further comprises a product image database, configured to store images captured by the imaging module. 
     
     
         20 . The system of  claim 11 , wherein the set of executable instructions, are configured, when executed, to cause the processor to associate and/or retrieve the classifier based on:
 a. product's packaging shape or shape descriptor, wherein the shape or shape-descriptor is identified as the key-word on the package that is semantically related to the inserted product characteristic;   b. matching representative key-words of the inserted products, to key-words associated to the classifiers in the classifiers database; and   c. matching the classifier's associated descriptors in the classifier database to the inserted product descriptors by applying local image descriptors' matching algorithm.   
     
     
         21 . The system of  claim 1 , wherein the cart further comprises a transceiver in communication with the central processing module, the transceiver adapted to transmit and receive an ultra-wideband radio pulse configured to provide location of the cart and wherein the classifiers retrieved are dependent on the cart's location.