Premise
Women who have experienced domestic violence use art therapy to process these events. The purpose of this research is to explore two different approaches to image classification and to identify limitations.
Synopsis
Using technology to address a social problem as domestic violence has always been an interest of me. I often see in the news that women are coerced, beaten, or even killed by their own husbands. The direct way to call the police does not seem to be an option for many. Women who are traumatized probably cannot express themselves through words. Exactly how this type of violence is processed and identified, and how technology can help, was the goal of this research.
Introduction
One in three women worldwide experiences violence. Mostly are physical or sexual violence experienced by an intimate partner (The Shadow Pandemic: Violence against Women during COVID-19, n.d.). These are the numbers before the outbreak of COVID 19 in 2020. Since the outbreak, all types of violence against woman and girls, particularly domestic violence, has intensified. The increasing violence against women is called the “Shadow Pandemic” which is growing amidst the COVID-19 crises (The Shadow Pandemic: Violence against Women during COVID-19, n.d.).
In case of experiencing domestic violence, there are many opportunities to seek help. The victims can call the police, ask friends and family for help but can also use helplines. In Europe, 46 numbers exist to call in case of domestic violence (The EU Is Supporting Stakeholders Working to End Violence against Women and Girls at the EU, National and Grassroots Levels., n.d.). Recognizing the importance of digital communication, major crisis helpline organizations began to offer online help services via chat (instant messaging) and e-mail (Mokkenstorm et al., 2016).
Research
However, the author Elena Sharratt (2016) has recognized that the lived reality of traumatized women contradicts completely at odds with this description. She analyzed in a group workshop that traumatized women are not likely to talk about their experiences but share these with painting, drawing, and creative writing projects and display non-verbal expressions of emotion through crying (Sharratt, 2016). The danger of verbally expressing traumatic events is the inducing or “triggering” of more stressful memories and even vivid and sensory flashbacks (Sharratt, 2016).
While drawing is a kind of individual expression, it can also be a communicative tool. Drawing tends to recount far more things to the reader than language (Farokhi & Hashemi, 2011). Accordingly, drawing is identified as a symbolic method of communication that exposes the artist’s psychological involvement in the experience (Venkatesan & Peter, 2018). Creative arts therapy is based on the proposition that people can use creativity to elicit emotional expression through alternative means of communication and can help individuals explore their emotions and experiences without the risk of retraumatization (Ikonomopoulos et al., 2017).
Approach
During this project, my goal is to get a better understanding of image recognition and the images which are necessary for the model.
Process:
Through the literature has shown that especially children use art therapy to process trauma. For this reason, the focus of the first iteration is on pictures painted by children. The first step for this is to collect the images. The second step is to classify these images and measure the performance of the model. To be able to classify the images, I used Lobe. In the second iteration, the same steps were done with a different data set. In the last iteration, the image classification was coded with python.
First Iteration
1.Scraping Pictures of Children where the mother experiences violence
Searching the internet for images of children working through trauma was not easy. There are hardly any and the images are quite scary. I have put the focus on the pictures where the children draw that the mother experiences violence. Through various forums and posts on the Internet, I found a total of 20 pictures. I have scrapped these pictures.
2.Scraping Pictures of Children that draw a happy family
The model had to be able to classify whether domestic violence is experienced or not. For this, I also needed pictures where the children painted a happy family. I also found these online through forums and posts and scraped them as well. I also scraped a total of 20 pictures of a happy family.
3. Image classification with Lobe
After I had both datasets, I used Lobe to classify these images and measure performance. Lobe is a tool to train easy and free machine learning models.
To train the model with Lobe I selected 15 images of violence at home and 15 images of a happy family. The prediction was quite good 93% of the images were predicted correctly, 7% were predicted incorrectly. Afterward, I tested the model with new images. Therefore, I used some images from the internet and captured new images through the camera that testers had painted for me.
The outcome of the prediction was really good. The new images have achieved 100% accuracy.
These are interesting impressions but from this dataset, it was obvious that they experience violence at home. Also, there have been very few pictures. The pictures of the women painted during therapy are not so obvious. So, what might these drawings of the women look like? That was the biggest hurdle in this research – there are almost none. There were videos where you could scrape the images, but again that would have been too few images. For this reason, I chose a dataset from Kaggle to repeat these steps.
Second Iteration
In the second iteration, I used the following dataset of Kaggle: https://www.kaggle.com/thedownhill/art-images-drawings-painting-sculpture-engraving/code. Through research, I have realized that mainly painting as well as drawing is done during art therapy. For this reason, I have chosen this data set and used only the folders: drawings and paintings.
Let’s use these 2 folders to build a model which predicts drawings and paintings. The folders consist of training and validation datasets. In total there are about 3500 images.
Examples of the drawings:
Examples of paintings:
Thereby, I also used Lobe and measured the accuracy of the classification.
Therefore, I only chose 424 images in total because Lobe had issues with more images. The prediction is really good – the model can predict 99% of images correctly.
After training the model, the model is validated. The validation set was also by Kaggle.
The model is even after validating as well as it was before. In general, it can be concluded that Lobe has a very high accuracy in image classification, but it cannot capture many images.
Third Iteration
In the last iteration, I went on to create a Convolutional Neural Network (CNN) in Python. Therefore, I used Tensorflow Keras Library to build the model. In contrast to Lobe, I can use more images and am not limited to about 400 images. It was the first time I learned and applied machine learning myself. I have had no previous experience in this area. In the process, I went through many tutorials of DataCamp and YouTube and always came to my limits. With the help of others, I was able to do the code for image classification.
The images of Kaggle first cropped to the same format and then all reduced to the same size and converted into a dataset.
The training dataset performed better than the validation data. The training data are relatively constant in contrast to the validation data. However, the training accuracy never reaches 100%. The validation accuracy reaches an accuracy of 100% but is also getting lower than 60%.
In general, the accuracy of Lobe was better than my code. Though, the code is based on a larger data set. However, I had very big difficulties with the code since I am not the most experienced programmer. The next step would have been to optimize this model, but that did not work.
Conclusion
In general, it can be recorded that the image classification is well done. With the Lobe software, there was better accuracy, but the code with python was able to capture more data. I learned a lot about image classification and machine learning. In the last weeks, I was able to build up knowledge about Tensorflow and Keras as well as explore the field of domestic violence against women. However, the project will not be continued at this point. The reason is that the implementation of the actual idea with image recognition for women who have experienced domestic violence is not feasible in a meaningful way. The collection of the necessary images is difficult to realize and other limitations have arisen.
References
Farokhi, M., & Hashemi, M. (2011). The analysis of children’s drawings: social, emotional, physical, and psychological aspects. Procedia-Social and Behavioral Sciences, 30, 2219-2224.
Ikonomopoulos, J., Cavazos-Vela, J., Vela, P., Sanchez, M., Schmidt, C., & Catchings, C. V. (2017). Evaluating the effects of creative journal arts therapy for survivors of domestic violence. Journal of Creativity in Mental Health, 12(4), 496-512.
Mokkenstorm, J. K., Eikelenboom, M., Huisman, A., Wiebenga, J., Gilissen, R., Kerkhof, A. J., & Smit, J. H. (2017). Evaluation of the 113Online suicide prevention crisis chat service: outcomes, helper behaviors and comparison to telephone hotlines. Suicide and Life‐Threatening Behavior, 47(3), 282-296.
Sharratt, E. (2016). Group Narratives of Trauma and Healing: Community Storytelling as a Critique of Individual ‘Talking Therapy’amongst Survivors of Sexual Violence. In Narrating Illness: Prospects and Constraints (pp. 75-82). Brill.
The EU is supporting stakeholders working to end violence against women and girls at the EU, national and grassroots levels. (n.d.). The European Commission. https://ec.europa.eu/justice/saynostopvaw/helpline.html
The Shadow Pandemic: Violence against women during COVID-19. (n.d.). UN Women. https://www.unwomen.org/en/news/in-focus/in-focus-gender-equality-in-covid-19-response/violence-against-women-during-covid-19
Venkatesan, S., & Peter, A. M. (2018). ‘I Want to Live, I Want to Draw’: The Poetics of Drawing and Graphic Medicine. Journal of Creative Communications, 13(2), 104-116.