Development of an Image Recognition Model Using an Image Search Function Based on Multiple Pre-Trained Models

Machine learning is widely utilized for data anal- ysis and decision-making, with supervised and unsupervised learning being the primary approaches. However, models suffer from overfitting, where they become overly adapted to the training data. To address this issue, semi-supervised learning has been employed. Semi-supervised learning is an effective technique for dealing with large datasets that are difficult to label, but it faces limitations in fields where ensuring data diversity and quantity is challenging. This paper proposes a robust image recognition model utilizing image search func- tions from Google. The proposed model improves accuracy by utilizing the order of search results to collect a variety of data and evaluating their reliability. In this paper, the order of search results is defined as ”image search depth” to measure the correlation between reliability and accuracy. While it is easy to collect large amounts of data from Google through automated methods, there is a risk that unrelated data could be included, potentially affecting the model’s accuracy. To address this issue, the model is trained with automated preprocessing. As part of this preprocessing, inference is performed on all images in the dataset using multiple pre-trained models that were trained on randomly selected images from the dataset to compute predictions. Images with the prediction above a certain threshold are selected as training data to enhance the final model ’s accuracy. To assess the contribution of preprocessing to accuracy improvement, we calculate accuracy by varying the number of parallel pre-trained models and the threshold values. Furthermore, the final model is evaluated using CIFAR- 100 to objectively demonstrate its performance. The results indicate that image search depth does not contribute to model accuracy, while the number of parallel pre-trained models and the threshold significantly impact accuracy.