Week #7 — Identification of Artists by Their Paintings with Machine Learning



Original Source Here

Week #7 — Identification of Artists by Their Paintings with Machine Learning

Finally, we have achieved the end of our project. We have reached our final results by implementing attention into our algorithm.

We have decided to implement Channel-wise and Spatial together according to related works. They provide informative content and the region.

Attention on Paintings

Spatial Attention
The spatial attention works as a complement to channel attention, focusing on the informative ‘where’ part. As an application, we made the kernel size 7 first. After the channel-wise operations, we applied a layer to obtain a spatial map.

In this way, desired regions will be highlighted while others will be suppressed. The spatial attention is calculated by using sigmoid function.

Channel-wise Attention
Channel attention focuses more on the meaningful parts of the images. For implementation, the spatial attention features are squeezed to make ineffective at first. Then we combined average pooling and max pooling on a shared layer.

After taking the totals of the results, we used the sigmoid as the activation function again.

Results

CNN Metric Results Before Attention (ResNet-152 + Fine-Tuning)
CNN Metric Results After Attention

The metric results after Atttention are shown in table. After Attention, we observed an improvement in results in general and a slight more increase in Van Gogh’s results in particular. We easily observed the effect of the methods we used on the results. We found that Attention and Fine-Tuning were effective. But still, we were not satisfied with Picasso’sresults and we thought about it. We observed the most misclassifications in Pablo Picasso’spaintings. We think this may be due to the different styles of his paintings that make it difficult for our model to learn. In the figure as shown below, nine of Picasso’s incorrectly predicted works are shown. As you can see, there is a serious style difference between his works. We think that this may prevent us from achieving our goal

Conclusion

According to our results, we observed that these famous painters have a style that distinguishes them from each other and from other normal painters. We’ve seen their unique style make them up. We observed that every step in our algorithm (CNN, Fine-Tuning, Attention) % style and model relationship,
that we applied to prove this hypothesis positively supports us to approach the result. Except for Picasso, we achieved the accuracy we wanted in other artists. We got better results than we expected, especially in Albrecht Dürer’s works. When we look at Albrecht Dürer’s works, we see that his works are generally in black and white, so we think that his style is easier to predict than others. However, we think by looking at Picasso’s works that the variety of styles in the artist’s works makes it difficult for us to reach our goal. We think the artist’s variety of styles makes it difficult for our model to learn. We can say that the most effective thing for both artists is their style. The monotony or diversity of their styles is affecting our algorithm. Thus, we observe the importance of styles in identifying artists.

Future Work

We think that better results can be achieved by examining the features such as the art movements used by the artists, the times they lived, and their nationalities in future works.

In addition, we think that different styles and movements can be observed and different inferences can be achieved by trying different artists and/or increasing the dataset.

The End…

We have come to the end of another semester. Thank you for your interest in our project. See you in the next projects. 🙂

References

https://openaccess.thecvf.com/content_ECCV_2018/papers/Sanghyun_Woo_Convolutional_Block_Attention_ECCV_2018_paper.pdf

AI/ML

Trending AI/ML Article Identified & Digested via Granola by Ramsey Elbasheer; a Machine-Driven RSS Bot

%d bloggers like this: