pattern-recognition

Project to advances in pattern recognition and tecnical experiences

Principal Results: Metrics choosed of classification: I choosed to use as classification metrics, the [accuracy, precision, f1-score, recall, precision], which has the characteristic:

• accuracy: Describes how the model performs across all the classes
• precision: The precision measures the model's accuracy in classifying a sample as positive.
• recall: Quantifies the number of correct positive predictions made out of all positive predictions
• f1-score: Combines the precision and recall, and give a harmonic mean of them

This metrics are presented in each jupyter notebook.

1. Using the Bag of Visual Words: The results in this approach showed that the number of visual words is very important. When I utilized a small number, like 2 or 5, the confusion matrix dont´t showed convergency. As the value was increased, an increase in performance was seen, with a lot of computational cost, but the models takes a long time to converge, having horrible results in this approach.

1.1 Possibilities to improve the results:

• Define the type of load the dataset, without uses append in the images, dividing in batches sizes
• Uses methods of data augumentation to improve the quantity of images
• Search another metods to extract the quantity of visual words, beacuase kmeans is very slowly
• Compare the extract of keyponitns using SIFT and HOG

Due to this, feature extraction with vgg16 showed much better results.

2. Using the vgg16 cnn as features exctator: The results of this case are great. Each class of dataset has a good metric

2.1 Possibilities to improve the results:

• Use more elaborated methods of data augmentation
• Alterate the space collors os images dataset
• Test more variations of hiperparameters
• Test anothers CNNs as feature extractor and compare

---

How to run

1. CLone the project

2. Install the poetry

3. Execute: poetry install poetry shell

4. Now, you can open the jupyter notebook and execute