Classifying Raw Materials with Discriminative Illumination

Classifying raw, unpainted materials --- metal, plastic, ceramic, fabric, etc. --- is an important yet challenging task for computer vision. Previous works measure subsets of surface spectral reflectance as features for classification. However, acquiring the full surface reflectance is time-consuming and error-prone. In this project, we propose to use coded illumination to directly measure discriminative features for material classification. Optimal illumination patterns---which we call ``discriminative illumination"---are learned from training samples, after projecting to which, the spectral reflectance of different materials are maximally separated. This projection is automatically realized by the integration of incident light for surface reflection. While a single discriminative illumination is capable of linear, two-class classification, we show that multiple discriminative illuminations can be used for nonlinear and multi-class classification. We also show theoretically the proposed method has higher signal-to-noise ratio than previous methods due to light multiplexing. Finally, we construct a LED-based multi-spectral dome and use the discriminative illumination method for classifying a variety of raw materials, including metal (aluminum, alloy, steel, stainless steel, brass and copper), plastic, ceramic, fabric and wood. Experimental results demonstrate the effectiveness of the proposed method.


Jinwei Gu and Chao Liu. Discrminative Illumination: Per-Pixel Classification of Raw Materials based Optimal Projections of Spectral BRDFs. CVPR 2012.

Jinwei Gu and Chao Liu.Supplementary Document (with proof and other experimental details).


  A LED-based Multispectral Dome Light:

We design and build a LED-based multi-spectral dome light for classifying raw materials based on optimal projections of their spectral BRDFs. The dome has 25 LED clusters. Each LED cluster has six color LEDs which can be weighted individually to create a desired spectrum. They are driven with Arduino boards with PWM controls. We learn optimal illumination patterns from training samples, after projecting to which the spectral BRDFs of different materials can be maximally separated.

  Database of Raw Materials:

We focus on unpainted, raw materials which are classified based on their surface spectral BRDFs. The database includes metal, plastic, fabric, ceramic, and wood. Within the class of metal, we have samples of alloy (4130), aluminum (5052, 6061, 2024, 7075), steel (cold roll and hot roll), stainless steel, brass, and copper. In total, there are 100 sample plates.

  Discriminative illumination as a physically-based linear classifier:

Our core idea is to use coded illumination as a classifier. (a) A schematic diagram in which coded illumination acts as a linear classifier, after projecting to which the spectral BRDFs of different materials are maximally separated. (b) An example of aluminum-vs-alloy classification. The image is captured by one of the 150 LEDs of the dome which yields the best classification performance on training data. Its classification rate on testing data is 41\%. (c) We train a linear kernel SVM classifier from the same training data, with the classification rate of 95% on the testing data. The bar graph shows the learned SVM light, w, where the 25 bar groups correspond to the 25 LED clusters and the six bars within each group correspond to the six LEDs. The vertical axis shows the relative brightness of each LED. Since the SVM light, w, has negative values, we implement it as the difference of two nonnegative vectors. (d) and (e) show the corresponding light patterns on the top view of the LED dome. The colors of the nodes show the spectra of the LED clusters. (f) and (g) show the corresponding captured images. (h) shows the difference of (f) and (g), which is used for classification. (i) is the classification result, shown as a binary image.

  Extension to multi-class classification:

Multiple discriminative illumination can be used for multi-class classification. We show an example of fabric-vs-ceramic-vs-plastic classification using the one-vs-all strategy. (b) The captured image under one of the 150 LEDs. (c) The learned three SVM light vectors. (d) To handle negative values, we implement the three SVM lights as four nonnegative light patterns. (e) shows one of the four captured images under the SVM lights. (f) the classification result. The classification rate is 94%. (g) In comparison, if we only select three LEDs for classification, we can at most have 62% classification rate.

  From linear to nonlinear classifier using multiple discriminative illuminations:

(a) A cascade classifier for the detection problem, which minimizes false positive rate by adding more stages while maintaining a small given false negative rate for each stage. (b) A toy nonlinear example of detecting red + from blue circles. (c) Classification results of the cascade classifier in which each stage is a linear classifier. Top: the classification results on the training samples. {Bottom: the classification boundaries.

  Aluminum detection using a cascade classifier:

We train a four-stage cascade classifier to detect aluminum from three other materials (ie, steel, ceramic, plastic). (a) captured image under one of the 150 LEDs. (b)(c)(d)(e) show the learned classifier and corresponding classification result (as a binary image) for each stage. A linear-kernel SVM classifier is used for each stage. With the false negative rate for each stage to be 2%, the four-stage cascade classifier has false negative rate of 4.2% and false positive rate of 0.07%. In comparison, if we train a single linear light for this problem with the same false negative rate, the false positive rate is 6%.

  Deal with surface normal variation:

We deal with surface normal variations by augmenting the training datasets with rotations of BRDF vectors. This works for small variations (within +/- 10 degrees). This picture shows a simulation result. (a)(b) Renderings of two BRDFs under natural lighting. (c) A sample with random surface normal ($\pm 10$ degrees variation in the tilt angle, color coded). (d) The distribution of the two BRDFs on the sample. (e) Measured image under a point light, with which it is difficult to separate the two BRDFs. (f) Measured image under a discriminative illumination, with which we can separate the two BRDFs more accurately.


  Supplementary Video:

This video show the dome light in action. Here are several more videos:


CVPR 2012 Oral Presentation (with videos)

ICCP 2012 Poster (PDF)