Item Recyclability Identification Based on Image Classification

Xiaoyi Chen

Background

When people are throwing away something seemingly useless, they are often in doubt: Is this recyclable or not? Recyclable items are valuable material rather than trash. By reusing recyclable items, a huge amount of energy cost to manufacture new materials and to deal with waste can be saved to help sustain a greener living environment for us all. Many kinds of glass, paper, metal, plastic, textiles, and electronics are recyclable materials, and should not be put into trash bins where non-recyclable materials should stay. [1] They should be thrown into recycling bins and later brought to a collection center and reprocessed into brand new materials. But can people memorize all that recycling knowledge? Do they bother to bring a recyclopedia manual with them? In most cases, the answer is unfortunately "No". Most people today, however, won't bother to bring a smartphone wherever they are going. Smart phones have become mighty portable computing devices with touch control and cameras. Smart phones can take images of items that will be thrown away can be processed by computer vision and classified by supervised machine learning methods, thus intelligently tell users the recyclability of their items without pain.

Method

Depending on the training image datasets, I will decide whether computer vision techniques such as edge detection, distance transform, skeleton extraction, contour extraction will be applied to extract the major features of the items. Principal component analysis might be employed to reduce the dimensionality then.
I plan to investigate on several classification methods on their accuracy and performance to decide which one is most suitable for this task.

SVM-KNN method has been proved effective for visual category classification. [2]
Gaussian Processing can provide value of confidence for classification. [3]
Metrics Learning can significantly improve the accuracy of KNN classification. [4]
Random Forest is less computational expensive than general classification methods. [5]

Datasets

There is no public datasets of images appropriate for deciding the recyclability of an item. I plan to collect hundreds of images to build a dataset. An image in the dataset contains one item, either recyclable or non recyclable, and is properly labeled. Datasets will be enforced the following restrictions to improve the accuracy of identification.

Image should be taken in a clear white blank background to ensure the item does not confuse with irrelevant objects.
Item must be entirely contained within the image.
Item should be placed at a fixed pose. This restriction is possibly not necessary if we use techniques like class segments sets [6] to make classification results invariant to planar transformation, especially rotation.

Timeline

Apr 12: Project Proposal
Apr 20: Training Image datasets collected. Investigate further on classification methods.
Apr 27: Dataset preprocessed by Computer Vision methods.
May 10: Project Milestone. Classification method implemented.
May 20: More classification method implemented.
May 25: Poster prepared. System tuned for better accuracy and performance.
May 31: Final Writeup.

References

[1]http://en.wikipedia.org/wiki/Recycling

[2]Hao Zhang, A.C. Berg, M. Maire, J. Malik. SVM-KNN: Discriminative Nearest Neighbor Classification for Visual Category Recognition. Computer Vision and Pattern Recognition IEEE Computer Society Conference. pages 2126-2136. 2006.

[3]Rasmussen, Carl. Gaussian Processes in Machine Learning. Advanced Lectures in Machine Learning. vol 3176. pages 63-71, 2004. Springer Berlin.

[4]Lorenzo Torresani, Kuang-chih Lee. Large Margin Component Analysis. Advances in Neural Information Systems. pages 1385-1392. Cambridge, MA, 2007. MIT Press.

[5]A. Liaw and M. Wiener. Classification and regression by randomForest. R News, 2/3:18–22, December 2002

[6]Kang B. Sun, Boaz J. Super, "Classification of Contour Shapes Using Class Segment Sets," cvpr, vol. 2, pp.727-733, 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05) - Volume 2, 2005