machine vision
Further work - identification algorithm
Slide 34: Conclusions
Further work - feature generation
Slide 33: Conclusions
image pre-processing may be improved by using more effective combinations of smoothing and gradient operators. Different texture analysis methods should also be tested. If needed, code optimisation and data sampling could dramatically improve the speed of the algorithm. However, image processing times are at the moment very reasonable. The feature extraction process should also be further investigated, testing different extraction masks and attempting selection of the extracted features.
Conclusions
Slide 32: Conclusions
summary of results
the proposed method produces high performing MLP solutions. The small magnitude of the standard deviations shows the robustness of the approach. Particularly important is the low rate of detection of false positives. The proposed algorithm gives comparable performances to other more complex algorithms in the literature.
learning results - visual comparison
Slide 30: Experimental results
plot of the average accuracy within the span of plus or minus one standard deviation.
Experimental results
Slide 29: Experimental results
learning results obtained using the three feature extraction masks. For each feature vector, two columns refer to the average accuracy results and their standard deviations obtained using the best MLP configuration respectively with one and two hidden layers. The table also breaks down the classification accuracy on each of the two classes and reports the number of BP training cycles.The estimates refer to 10 independent learning trials.
Training and validation sets
Slide 28: Proposed approach
ANN identification
these are the main features and training parameters of the proposed classifier.
Feature extraction masks
Slide 26: Proposed approach
These are the 3 masks that were used in the experiments. They generate respectively a 25-dimensional, a 49-dimensional and a 81-dimensional feature vector of temperature gradient angles.
Feature extraction visual example
Slide 26: Proposed approach
this figure visualises the feature extraction process. Individual cells of the SST map grid do not correspond to actual pixels but are only for illustration purposes. Gradient directions are also only indicative.
