Hazmat Label Identifier

The hazmat label identifier was designed to detect a defined range of hazmat labels in a range of environments. This was done through identifying a range of features each label posessed and return these to the user. It was not in the scope of this project to classify to the label itself. The range of labels the system needed to be able to classify various text, graphics and colours can be seen below:

The development of determining what these attributes were or whether they existed at all required the use of OpenCV and basic python modules. The key functional requirements were the ability to segment and orientate a label from the environment as well as segment and classify the attributes such as the top and bottom colour of the label, class number which appears in the bottom corner, text and label identification.

An example output of the 5 key attributes can be seen below:

top: red
bottom: red
class: 3
text: COMBUSTIBLE
symbol: Flame

The key design decisions include:

1. Segmenting the label from the background

This was achieved through a range of adaptive thresholding and edge detection processes given the strong edges the label possessed. This was then combined with some contour processes to ensure it matched properties that one would associate with the shape of the label.

2. Colour Determination

Colour recognition was fairly 'dumb' and used the HSL colour space with pre-defined colour ranges to determine how many pixels were in a given pre-defined colour range.

3. Text Recognition

This process involved segmenting each character and then determing the histogram of oriented gradients of scaled shape which was then passed into an SVM where the expected character was output. This was done for each segmented character and a dictionary was used to look up words that may be expected. If their was a match between a dictionary word and the OCR's prediction, it was output as the text field.

4. Label Recognition

Label recognition was done in the same manner as text recognition, with some minor differences in the segmentation approach being somewhat simpler.

5. Class Number Detection

Similar to the text and label recognition, an SVM was used to classify a HOG descriptor The one difference was these characters were segmented using blobbing which returns a far cleaner outline. For this reason the resulting performance was signifigantly better.

Results

The identifier was effective under a range of backgrounds and conditions, but struggled were lighting conditions were not optimal. Below are some results from a test set:

Image	Output
	top: yellow bottom: yellow class: 5.1 text: OXIDIZER symbol: Oxidiser
	top: yellow bottom: white class: 7 text: RADIOACTIVE III symbol: Radioactive
	top: white bottom: black class: 8 text: None symbol: Corrosive
	top: red bottom: red class: 3 text: COMBUSTIBLE symbol: Flame
	top: white bottom: white class: 2 text: INHALATION HAZARD symbol: Skull and crossbones on black diamond