ImageSearch

This is demo to upload image and create image vector in the neo4j database .

Later vector is used for image search

---

Flow :

---

We convert images to vectors using TensorFlow and store these vectors as properties in Neo4j.

By utilizing Neo4j's vector index feature, we can run cosine similarity algorithms on these vectors to find similarities.

Additionally, a website built on Flask can be used to upload images and perform searches based on these vectors.

---

There is different script for downloading images from internet and upload to neo4j

---

Configration in Neo4j

---

Neo4j version : 5.20.0

Install APOC : apoc-5.20.0-core.jar

Create vector index

CALL db.index.vector.createNodeIndex('imageCosine', 'Image', 'vectorCosine', 2048, 'cosine');

CALL db.index.vector.createNodeIndex('imageEuclidean', 'Image', 'vectorEuclidean', 2048, 'euclidean'); // Optional 

---

Configration for Flask website

---

1. pip install -r requirment.txt
2. export FLASK_APP=app.py
3. Edit ImageSearchNeo4j.py (line 12) with correct neo4j connection string
4. flask run --port 5002

driver = GraphDatabase.driver("neo4j+s://auraID.databases.neo4j.io", auth=("neo4j", "auraIDpassword"))

---

Bulk download and upload of images

---

To download use : ImageSearch/bulkUpload/upload/download.py

This will read images_full.csv and download as per that .

To get csv : https://storage.googleapis.com/openimages/2016_08/images_2016_08_v5.tar.gz

To upload to Neo4j : ImageSearch/bulkUpload/UploadToNeo.py

This will upload images to neo4j . Edit the connection string for neo4j connection

---

For better performance to create image vector install CUDS drivers

https://developer.nvidia.com/cuda-downloads

---

Cypher to upload image vector

CREATE (I:Image {name: $name,desc:$desc})-[:Original]->(:Originalbase64 {base64:$base64})
CREATE (I)-[:Compress]-> (:Compress{base64:$compressed_base64_str})  
WITH I 
CALL db.create.setNodeVectorProperty(I, "vectorCosine", $vector)

---

Cypher to search image vector

CALL db.index.vector.queryNodes('imageCosine', 5, $vector)
YIELD node AS similarAbstract, score
RETURN  similarAbstract.name,score