Palm Recognition

Goal:

The objective of the project is to develop a robust model that can extract features from palm images and maximize the distance between inter-class while minimizing intra-class distances. We will begin by using a dataset with 370 identities to train a classification model for feature extraction. During the inference phase, we will select N images from each identity for registration and use the remaining images as query images for recognition. For accurate recognition, we will pair the query images with registered IDs using cosine similarity.

Dataset:

11KHands: The 11k Hands dataset is a collection of 11,076 hand images of 190 individuals, ranging in age from 18 to 75 years old. The dataset includes images of both the left and right hands, with each hand being photographed from both the dorsal and palmar sides against a white background. The images are accompanied by metadata, which includes information such as the subject's ID, gender, age, and skin color, as well as details about the hand image itself, such as which hand it is, which side was photographed (dorsal or palmar), and whether the image contains accessories, nail polish, or irregularities.

Prerprocess:

The goal of the project is to identify a person through the palmar, so we remove the dorsal data and only focus on the palmar hand-side. The configurations for preprocess can be found in params.yaml.The project use the open source library for preprocessing:

1. rembg: background removal
2. mediapipe: hand keypoint detection

python src/preprocess.py

Train:

The configurations for train can be found in params.yaml

python src/main.py

Convert the model to ONNX:

The configurations for onnx can be found in params.yaml

python src/to_onnx.py

Inference Example with ONNXRUNTIME:

python src/onnx_inference_workflow.py

Register(Gallery) and Query(Probe) Similarity Search:

The project uses Facebook FAISS for similarity search. FAISS is a library for efficient similarity search and clustering of dense vectors, developed by Facebook AI Research (FAIR). It provides a range of search methods, including exact search, approximate search, and k-nearest neighbor search, which can be used in a variety of applications such as image and text search, recommendation systems, and data compression. FAISS is highly optimized for large-scale datasets and is capable of handling billions of vectors on a single GPU. It is open-source and available for use by researchers and developers.