Method and system for product recognition

1. A method of product recognition, comprising

decomposing a product image into a set of frequency components;

performing a frequency analysis for each frequency component in the set of frequency components;

detecting a plurality of edge pixels of the image based on the frequency analysis;

selecting a set of interested pixels in the region of each of the plurality of edge pixels;

creating a quantized product image; and

creating a normalized histogram of the quantized product image using the sets of interested pixels selected for each of the plurality of edge pixels, wherein product recognition is based on similarities between the normalized histogram of the quantized product image and images of similar products in a database of products.

2. The method of claim 1 wherein performing a frequency analysis comprises obtaining local intensity thresholds and spatial frequencies for each frequency component in the set of frequency components.

3. The method of claim 2 wherein decomposing the product image comprises:

converting the quantized product image to a Hue, Saturation and Value colour space;

extracting a Value channel of the Hue, Saturation and Value colour space;

applying a derivative of Gaussian filter to the extracted Value channel; and

decomposing the images derived from the application of the derivative of Gaussian filter in a horizontal and vertical orientation into a first intrinsic mode function representative of the horizontal orientation and a second intrinsic mode function representative of the vertical orientation of the Value channel.

4. The method according to claim 3 wherein obtaining spatial frequencies for each frequency component further comprises estimating a spatial frequency for each pixel associated with each intrinsic mode function and wherein detecting the plurality of edge pixels is based either on local intensity thresholds or the estimated spatial frequencies.

5. The method of claim 3 wherein selecting the set of interested pixels comprises:

for each of the plurality of edge pixels:

applying a window having a constant size to each edge pixel to determine a set of pixels in the region of the edge pixel;

calculating an entropy associated with each of the pixels in the set of edge pixels;

selecting the set of interested pixels having maximum entropy.

6. The method of claim 5 wherein decomposing the images derived from the application of derivative of Gaussian filter further comprises applying Empirical Mode Decomposition, EMD, to a response image from derivative of Gaussian filter.

7. The method according to claim 5 wherein detecting the plurality of edge pixels of the image comprises detecting a plurality of extrema in each intrinsic mode function.

8. The method according claim 1 wherein creating the quantized product image comprises quantizing the product image using a set of perceptual colours.

9. The method according claim 3 wherein detecting the plurality of edge pixels further comprises pre-processing the value channel.

10. The method according to claim 9 further comprising classifying a plurality of edge pixels associated with the value channel into a plurality of foreground and background clusters.

11. The method according to claim 10 wherein classifying comprises applying Sauvola's thresholding to the value channel to extract a binary image.

12. The method according to claim 11 wherein a final image is estimated based on a distance value between the binary image and a background surface.

13. A non-transitory computer readable programmable medium carrying a computer programme stored thereon which when executed by a processor implements the method according to claim 1 .

14. An apparatus for product recognition, comprising

means for decomposing a product image of into a set of frequency components; means for performing a frequency analysis for each frequency component in the set of frequency components;

means for detecting a plurality of edge pixels of the image based on the frequency analysis;

means for selecting a set of interested pixels in the region of each of the plurality of edge pixels;

means for creating a quantized product image; and

means for creating a normalized histogram of the quantized product image using the sets of interested pixels selected for each of the plurality of edge pixels, wherein product recognition is based on similarities between the normalized histogram of the quantized product image and images of similar products in a database of products.

15. A system for product recognition comprising the apparatus according to claim 14 .

16. A method of product recognition, comprising

decomposing a product image into a set of frequency components, wherein decomposing the product image includes:

converting the product image to a Hue, Saturation and Value color space;

extracting a Value channel of the Hue, Saturation and Value color space;

applying a derivative of Gaussian filter to the extracted Value channel; and

decomposing the images derived from the application of the derivative of Gaussian filter in a horizontal and vertical orientation into a first intrinsic mode function representative of the horizontal orientation and a second intrinsic mode function representative of the vertical orientation of the Value channel;

performing a frequency analysis for each frequency component in the set of frequency components, wherein performing a frequency analysis comprises obtaining local intensity thresholds and spatial frequencies for each frequency component in the set of frequency components, wherein obtaining spatial frequencies for each frequency component further comprises estimating a spatial frequency for each pixel associated with each intrinsic mode function;

detecting a plurality of edge pixels of the image based on the frequency analysis, wherein detecting the plurality of edge pixels is based either on local intensity thresholds or the estimated spatial frequencies;

selecting a set of interested pixels in the region of each of the plurality of edge pixels, wherein selecting the set of interested pixels comprises:

for each of the plurality of edge pixels:

applying a window having a constant size to each edge pixel to determine a set of pixels in the region of the edge pixel;

calculating an entropy associated with each of the pixels in the set of edge pixels; and

selecting the set of interested pixels having maximum entropy;

creating a quantized product image; and

creating a normalized histogram of the quantized product image using the sets of interested pixels selected for each of the plurality of edge pixels, wherein product recognition is based on similarities between the normalized histogram of the quantized product image and images of similar products in a database of products.