Image Search System Based on Feature Vectors of Convolutional Neural Network

Edge computing offers real-time applications because the edge device closes with the data source such as the end device. This condition gives the challenge to implement deep learning in the edge device. Unfortunately, deep learning requires high computing resources, but often edge-side devices have limitations. In this study, we built an image search system based on CNN (Convolutional Neural Network)'s feature vectors to address the challenges by enlarging the implementation of CNN in the edge device such as Raspberry Pi 3. The image search system applied these informative features vector to get similar images in the image searching task by using cosine similarity. We used a 102-flower categories dataset and we prepared a light database to run the system as an off-line system in the edge device. The MobileNetV2 as CNN's model reached 70.02% of the top 1 accuracy and 92.84 % for the top 5 accuracy. As a result, the image search system showed five images result with the most similar image from the same image category. Image resolution, model complexity, and hardware capability give the significant time in this image search system. The framework of this system can be simply used for other deep learning models and applications by updating the model, dataset, database, and hardware.