Media images significantly shape our perception of reality and hold considerable power in constructing and reinforcing social narratives, particularly concerning underrepresented groups like migrants. Unlike textual communication, visual media relies on association, triggering more potent emotional responses and potentially leading to insidious stereotyping. Numerous studies have demonstrated the link between biased media portrayals and discriminatory actions against minority groups. This research specifically focuses on the visual representation of migrants, addressing the lack of large-scale, cross-national studies utilizing advanced analytical techniques. Prior research has often examined limited media samples or focused on textual analysis. This study leverages the power of deep learning to analyze a substantial volume of visual data from ten countries, offering a comprehensive analysis of the visual representation of migrants, refugees, and expats across various demographic and emotional dimensions. The chosen countries represent a spectrum of attitudes toward migrants, allowing for cross-cultural comparisons.

Existing literature extensively examines media representations of migrants, highlighting the prevalence of negative stereotypes and the association of migrants with threats to national security or social welfare systems. Studies reveal an under-representation of migrant women and a tendency to conflate migrants with refugees or asylum seekers, often portraying them in dehumanizing ways such as massified crowds. The concept of the "ideal refugee" has also been discussed, contrasting with the often-negative portrayals of real refugees. Disparities in media coverage exist based on the political leanings of media outlets, with left-leaning sources tending to portray migrants as victims and right-leaning sources framing them as a threat. The terms "immigrant," "refugee," and "expat" are shown to be politically charged and context-dependent, reinforcing existing power dynamics and often masking the diversity within migrant populations. While several studies focused on individual countries or specific media outlets, this study attempts a more comprehensive, cross-national investigation using advanced computational methods.

This interdisciplinary study employed a mixed-methods approach combining deep learning techniques with social science perspectives. Images associated with "migrants," "refugees," and "expats" were collected from Google's Custom Search API in ten countries, spanning the range of migrant acceptance levels according to the Gallup Migration Acceptance Index. A total of 17,898 images were collected, with a focus on images containing faces. Deep learning models were employed to analyze various visual aspects, including: 1. **Face and People Detection:** The Multi-task Cascaded Convolutional Networks framework was used to detect and extract faces. When the resolution was low, a super-resolution technique was applied to enhance the image quality. A separate model was used to detect people, even when faces were not clearly visible. 2. **Demographic Information:** The FairFace model was used to estimate gender, age, and facial features (White, Black, Latino Hispanic, East Asian, South-East Asian, Indian, and Middle Eastern). Rather than selecting the single most likely class, an average probability distribution was calculated for each image to mitigate biases inherent in the model. 3. **Emotional Information:** The EmoNet model was used to estimate the valence (positive/negative) and arousal (calm/excited) of emotional expressions in the faces. The model also provided a probability distribution over different emotion categories (Happy, Sad, Surprise, Fear, Disgust, Anger, Contempt, Neutral). 4. **Crowd Detection:** A pre-trained neural network and normalized squared Euclidean distance were used to calculate the probability of an image containing a crowd. This was done to analyze the tendency to dehumanize migrants by representing them as indistinguishable masses. The results were compared with official immigration statistics from various sources to validate the findings and understand potential biases. The Kullback-Leibler divergence was employed to quantify differences between distributions across groups and locations.

The study revealed several significant biases in the visual portrayal of migrants: 1. **Gender:** A significant underrepresentation of women across all migrant groups was observed, despite official statistics indicating a more balanced gender distribution in many migrant populations. This aligns with previous findings on the invisibility of migrant women in media, especially skilled female migrants. The discrepancy is particularly striking for expats, where women are significantly underrepresented. 2. **Facial Features:** Expats are predominantly depicted as "white," while migrants and refugees show a higher representation of "Black" and "Middle Eastern" features. This underscores the bias of associating skilled migration with whiteness. Comparison with US statistics revealed a considerable underrepresentation of "Asian" expats compared to official data on highly skilled migrants. 3. **Age:** Men were shown to be portrayed as older, and women as younger, than indicated by official statistics. Children were underrepresented across all groups, with the lowest proportion observed in images of expat women, suggesting a focus on career-oriented portrayals, thereby ignoring family aspects of expat life. 4. **Emotions:** Migrants and refugees were associated with negative emotions, while expats showed predominantly positive or neutral expressions. Significant gender differences were observed, with women more likely to display happiness and men more likely to show anger. These findings correlate with existing emotional stereotypes about gender and socioeconomic status, linking migrants with poverty and expats with wealth. 5. **Crowds:** Migrants and refugees were significantly more likely to be depicted in large, unidentified crowds, reinforcing the dehumanization of these groups. Expats, conversely, were more frequently shown as individuals or in small groups. 6. **Cross-National Differences:** Substantial variations in portrayal were observed across the ten countries studied, indicating context-dependent biases. Countries with higher migrant acceptance tended to show more balanced gender representation. Divergence in the portrayal of emotions was most pronounced between Sweden and Hungary, possibly reflecting differences in their migration policies and public discourse.

The findings of this study address the research question by demonstrating significant biases in the visual representation of migrants in media across ten countries. The mismatch between visual portrayals and official statistics highlights how media reinforces stereotypical narratives, often focusing on negative aspects of migration while neglecting the diversity of migrant experiences. The underrepresentation of women and children reinforces gendered and ageist stereotypes. The association of specific racial features with different migrant categories highlights underlying power dynamics and colonial perspectives. The analysis of emotions further exposes how media utilizes emotional cues to shape public perception. The variations observed across countries demonstrate the context-dependent nature of these biases, suggesting the influence of national migration policies, cultural norms, and political ideologies on media representation. This research contributes to a deeper understanding of how media imagery contributes to the construction of social identities and reinforces existing inequalities.

This study demonstrates the effectiveness of combining deep learning with social science methods to analyze large-scale visual data on migration. The findings reveal pervasive biases in media representations across various demographic and emotional dimensions. Future research could explore the causal links between these biases and their impact on attitudes toward migrants and policy-making. Further analysis of the role of specific media outlets and political contexts in shaping these representations is also warranted. The methods employed in this study could be applied to other social groups or types of media to examine similar biases.

The study is limited by the inherent biases present in the deep learning models used, particularly concerning age and emotion estimation. Although measures were taken to mitigate these biases, residual effects remain possible. The use of Google Images as a data source introduces limitations related to the diversity and representativeness of the image set, and the focus on images with faces might exclude relevant visual elements. The reliance on pre-trained models limits the ability to fine-tune the analysis to specific cultural nuances. Further research should address these limitations and explore alternative approaches for data collection and model training.