Efficient retrieval of a target from an image in a collection of remotely sensed data

Patent metadata

Field	Value
Publication number	US-11978236-B2
Application number	US-202117513011-A
Country	US
Kind code	B2
Filing date	Oct 28, 2021
Priority date	Feb 25, 2021
Publication date	May 7, 2024
Grant date	May 7, 2024

Abstract

State of art techniques performing image labeling of remotely sensed data are computation intensive, consume time and resources. A method and system for efficient retrieval of a target in an image in a collection of remotely sensed data is disclosed. Image scanning is performed efficiently, wherein only a small percentage of pixels from the entire image are scanned to identify the target. One or more samples are intelligently identified based on sample selection criteria and are scanned for detecting presence of the target based on cumulative evidence score Plurality of sampling approaches comprising active sampling, distributed sampling and hybrid sampling are disclosed that either detect and localize the target or perform image labeling indicating only presence of the target.

First claim

What is claimed is:

1. A processor implemented method for efficient retrieval of a target from an image in a collection of remotely sensed data, the method comprising:
receiving, by one or more hardware processors, the image from a set of images in the collection of remotely sensed data to retrieve the target from the image, wherein the target is defined by spatial features, spectral features, or a combination thereof;
selecting, by the one or more hardware processors, a first sample set of pixels from the image by applying a sampling technique, wherein the sampling technique comprises one of:
a) an active sampling technique that selects the first sample set of pixels based on location criteria, and successive sample sets in vicinity of a sample set identified for a previous iteration, wherein the location criteria comprises one of i) selecting pixels within a predefined radius around a pixel at center of the image, and ii) selecting the pixels within the predefined radius at a region in the image wherein pixels features of the region have highest reciprocal difference with respect to the pixel features associated with the target, and wherein the active sampling technique enables simultaneous detecting and localizing of the plurality of desired properties in the image; and
b) an evenly distributed sampling technique that selects i) the first sample set of pixels comprising evenly distributed pixels across the image, and ii) the successive sample sets by varying frequency of the evenly distributed pixels to gradually reduce spatial distance between the evenly distributed pixels, wherein the evenly distributed sampling technique enables detecting of the plurality of desired properties in the image;
estimating an evidence score, by ML models implemented by the one or more hardware processors by processing the first sample set of pixels, wherein the evidence score is indicative of magnitude of presence of the target in the image;
identifying, by the one or more hardware processors, the image to have the target and labelling the image as image of interest if the estimated evidence score computed for the first sample set of pixels, is above an evidence threshold;
selecting, by the one or more hardware processors, successive sample sets in relation to the first sample set of pixels, for the successive sampling iterations, wherein the selection is in accordance with the sampling technique applied;
computing, by the ML models implemented by the one or more hardware processors, a cumulative evidence score, for a current sample set of pixels and a previous sample set of pixels;
identifying, by the one or more hardware processors, the image to have the target, retaining labelling of the image as the image of interest, and terminating the cumulative evidence score computation if the cumulative evidence score is above the evidence threshold;
continuing, by the one or more hardware processors, the cumulative evidence score computation for the successive sample sets of pixels, if the evidence score computed in current iteration is below or equal to the evidence threshold; and
discarding, by the one or more hardware processors, the image if post entire processing of the image, the cumulative evidence score is below the evidence threshold.

2. The processor implemented method of claim 1 , wherein the active sampling technique comprises:
identifying a first pixel location in the center of the image and identifying pixels within a predefined radius of the first pixel location as the first sample set of pixels, wherein the radius is determined based on scale and resolution of the set of images to be processed, and in context to a pixel size identified for the target;
computing the evidence score for the first sample set and labelling the image as the image of interest if the evidence score is above or equal the predefined evidence threshold;
creating a plurality of child sample sets of pixels in the vicinity of the first sample set of pixels, if the evidence score is below the evidence threshold;
selecting a second sample set of pixels from among the plurality of child sets in a direction where gradient of evidence for the target is maximum;
computing the cumulative evidence score for the first sample set of pixels and the second sample set of pixels and labelling the image as the image of interest, with first sample set and second sample identified as location of the target; and
iteratively performing successive second sample set of pixel selection and the cumulative evidence score computation if the evidence score in the previous iteration is below the evidence threshold until the entire image is scanned.

3. The processor implemented method of claim 1 , further comprising applying a hybrid sampling technique that utilizes combination of even distributed sampling technique and the active sampling technique, wherein the hybrid sampling technique enables simultaneous detecting and localizing of the plurality of desired properties in the image, the hybrid sampling technique comprising:
segmenting the image into a first set of plurality of quadrants;
applying the even sampling technique to identify a plurality of first samples of pixels around each of a plurality of pixels locations placed at equal pixels intervals in each of the first set of plurality of quadrants;
computing the evidence score for each of the plurality of first samples of pixels for each of the first set of plurality of quadrants;
selecting a first quadrant among the first set of plurality of quadrants, wherein the selected first quadrant has highest cumulative evidence score calculated by aggregating the evidence scores of each of the plurality of first samples of pixels lying in the first quadrant;
segmenting the first quadrant into a second set of plurality of quadrants; and
iteratively performing the steps of hybrid sampling technique on the second set of plurality of quadrants till the target is detected and localized.

4. The processor implemented method of claim 1 , wherein the ML models are trained for labelling the images by computing evidence score using training data comprising a set of sampled images selected in accordance with the sampling technique used by the ML model.

5. The processor implemented method of claim 1 , wherein the evidence score is computed by determining an association between the presence of the target and pixels in the image, wherein the association is identified using: a) Spatial distribution of the target across the image, b) associating between the target and other properties present in the image, and c) associating relation between features of a part of image with the entire image.

6. The processor implemented method of claim 1 , wherein a ratio of a size of the first sample set of pixels to the size of the image is below a predefined value.

7. A system for efficient retrieval of a target from an image in a collection of remotely sensed data, the system comprising:
a memory storing instructions;
one or more Input/Output (I/O) interfaces; and
one or more hardware processors coupled to the memory via the one or more I/O interfaces, wherein the one or more hardware processors are configured by the instructions to:
receive the image from a set of images in the collection of remotely sensed data to retrieve the target from the image, wherein the target is defined by spatial features, spectral features, or a combination thereof;
select a first sample set of pixels from the image by applying a sampling technique, wherein the sampling technique comprises one of:
a) an active sampling technique that selects the first sample set of pixels based on location criteria, and successive sample sets in vicinity of a sample set identified for a previous iteration, wherein the location criteria

Assignees

• Tata Consultancy Services Ltd
• Tata Consultancy Ltd Services

Inventors

• Deshpande Shailesh Shankar
• Purushothaman Balamuralidhar

Classifications

• G06V2201/07

  Target detection · CPC title

• G06V10/245 Primary

  by locating a pattern; Special marks for positioning · CPC title

• G06T7/11

  Region-based segmentation · CPC title

• G06V10/25

  Determination of region of interest [ROI] or a volume of interest [VOI] · CPC title

• G06V10/44

  Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components · CPC title