Result analysis for the further improvement

Our current preliminary TSR (Traffic Sing Recognition) implemenatation is based on the bianarized image which uses the dominating color as its background, thus black, and all other colors as the forground, white. After we apply the line receptors above on it to compute the crossing state to use as the input for the trainned ANN. However, the binarized image in fact does not contain the important color information and hence the preliminary test results also reflects this point. This problem can be vividly visualized by using the feature correlation graph (FCG) that I introduced at Choose proper image abstractions or features for sign recognition. At this time, we computed the correlation between the state vector applied to the training traffic sign pattern set. The line receptors applied to create the line crossing state vectors is shown below.

tspr_fullset_final_receptors.png

After, the correlation between the each traffic sign set using this line receptor state difference, which is the number of equal state, is computed as shown in below.

linereceptorcorrelation.png

As illustrated clearly in the FCG graph, the line receptors applied to the binarized images do not have enough distinctive features at a large number of traffic signs. This means that we need to introduce more traffic sign features to make distinctions between traffic signs and expected extension should be made in following features: