Reverse Image Search Engine for Garment Industry

The Internet provides an extensive quantity of data in the form of text and images. Search Engines like Google and Bing are intended to retrieve the most relevant answers for our textual queries. In search engines, queries are generally matched between text-to-text or text-to-image which excludes the entire domain of image-to-image matching. Searching for images based on text poses the problem of retrieving results primarily based on image description and tagging. The ideal solution to tackle the flaws of text-based image queries and achieve image-to-image matching is obtained by extracting features of the image and using these features to retrieve similar images. Extracting information from visual data and similarity matching fall under the domain of Content Based Image Retrieval (CBIR). CBIR analyzes image information using low level representation including color, shape, texture, and spatial representation of objects to set up feature vectors for image indexing. Deep Learning techniques like Convolutional Neural Networks (CNN) serve the purpose of fulfilling the feature extraction task. Retrieval methods are achieved based on the low-level features of an image and mathematical similarity metrics aid in determining the relevant images satisfying the query for similarity. However, low level features do not capture the idea behind the image and hence limiting the images to a specific domain ensures that accuracy and similarity are as expected. The proposed project is specific to the domain of garments in the fashion industry. The proposed project conducted a study on the use of different pre-trained CNN models for the purpose of image feature extraction and attempted to fine-tune the selected ResNet50 model for the custom dataset. The feature similarity matching was performed by selecting a similarity metric famously used by music streaming giant Spotify's algorithm, Annoy Indexing, optimized for time and accuracy.