System and method for symbol detection
An optical symbol detector has an input operative to receive a captured image frame. The detector includes a feature extractor engine coupled to the input to produce a downscaled descriptor map based on the captured image frame, the descriptor map including a convolutional map of blocks of the captured image frame. The detector further includes a coarse estimator engine coupled to an output of the feature extractor engine to produce a detection indication and localization information of an optical pattern indicative of a machine-readable symbol based on the downscaled descriptor map.
1 . An optical symbol detector, comprising:
an input operative to receive a captured image frame;
a feature extractor engine coupled to the input to produce a downscaled descriptor map based on the captured image frame, the downscaled descriptor map including a convolutional map of blocks of the captured image frame; and
a coarse estimator engine coupled to an output of the feature extractor engine to produce a detection indication and localization information of an optical pattern indicative of a machine-readable symbol based on the downscaled descriptor map, wherein the coarse estimator engine includes a trained decision tree cascaded with a trained neural network.
2 . The optical symbol detector of claim 1 , wherein the feature extractor engine is operative to:
partition the image frame into blocks; and
for each block:
compute a downscaled descriptor based on an S-transform operation to produce a vector representative of that block; and
compute a convolutional descriptor based on the vector of the block convolved with a kernel that includes vectors of neighboring blocks;
wherein the downscaled descriptor map comprises a set of convolutional descriptors of the blocks.
3 . The optical symbol detector of claim 2 , wherein the blocks are non-overlapping.
4 . The optical symbol detector of claim 2 , wherein the blocks are square.
5 . The optical symbol detector of claim 2 , wherein the blocks are 8×8 pixels.
6 . The optical symbol detector of claim 2 , wherein in the computation of the convolutional descriptor, for each block, the kernel includes vectors representing neighboring blocks along each side of that block.
7 . The optical symbol detector of claim 6 , wherein the kernel is 3×3 blocks.
8 . The optical symbol detector of claim 1 , wherein the feature extractor engine is arranged to computed multiple downscaled descriptors of the descriptor map in parallel and independently of one another.
9 . The optical symbol detector of claim 1 , wherein the feature extractor engine is further operative to perform contrast normalization of the convolutional map to produce a contrast-normalized feature map that is input to the coarse estimator engine.
10 . The optical symbol detector of claim 1 , wherein the coarse estimator engine includes a plurality of computation paths, each computation path corresponding to a detectable pattern, wherein each computation path includes a decision tree trained to detect the corresponding pattern, cascaded with a neural network trained to detect the corresponding pattern.
11 . The optical symbol detector of claim 10 , wherein the localization information includes location, horizontal vector, and vertical vector information.
12 . The optical symbol detector of claim 1 , wherein the trained neural network includes a set of output nodes trained to produce localization information of the corresponding detection pattern.
13 . The optical symbol detector of claim 1 , further comprising:
a validation and fine estimator engine having a first input coupled to an output of the coarse estimator engine, and a second input coupled to the input operative to receive the captured image frame, wherein the validation and fine estimator engine further includes:
a neural network trained to extract a portion of the captured image frame received via the second input that corresponds to the optical pattern based on the first input, and to determine whether that portion corresponds to a known pattern type.
14 . A method for operating an optical symbol detector, comprising:
receiving a captured image frame;
autonomously computationally producing a downscaled descriptor map based on the captured image frame, the downscaled descriptor map including a convolutional map of blocks of the captured image frame; and
autonomously computationally producing a detection indication and localization information of an optical pattern indicative of a machine-readable symbol based on the downscaled descriptor map;
wherein producing the downscaled descriptor map includes:
partitioning the image frame into blocks; and
for each block:
computing a downscaled descriptor based on an S-transform operation to produce a vector representative of that block; and
computing a convolutional descriptor based on the vector of the block convolved with a kernel that includes vectors of neighboring blocks;
wherein the downscaled descriptor map comprises a set of convolutional descriptors of the blocks.
15 . The method of claim 14 , wherein in producing the downscaled descriptor map, the blocks are non-overlapping and square.
16 . The method of claim 14 , wherein in computing the convolutional descriptor, for each block, the kernel includes vectors representing neighboring blocks along each side of that block.
17 . The method of claim 14 , wherein producing the downscaled descriptor map further includes performing contrast normalization of the convolutional map to produce a contrast-normalized feature map.
18 . The method of claim 14 , wherein producing the detection indication and localization information of the optical pattern includes processing a plurality of computation paths, each computation path corresponding to a detectable pattern, wherein each computation path includes a decision tree trained to detect the corresponding pattern, cascaded with a neural network trained to detect the corresponding pattern.
19 . The method of claim 14 , further comprising:
extracting a portion of the captured image frame that corresponds to the optical pattern and determining whether that portion corresponds to a known pattern type.