Full Code of dlau/mineye for AI

Repository: dlau/mineye
Branch: master
Commit: 486e01eb75f3
Files: 10
Total size: 14.8 KB

Directory structure:
gitextract_aagimesr/

├── .gitignore
├── README.md
├── img.py
├── package.json
├── pages/
│   ├── home.jsx
│   └── home.less
├── server.py
├── static/
│   ├── img/
│   │   └── .gitignore
│   └── js/
│       └── .gitignore
└── webpack.config.js

================================================
FILE CONTENTS
================================================

================================================
FILE: .gitignore
================================================
.DS_Store
node_modules
bower
*.pyc
static/img/bank/**/*.jpg
.idea
bank.db


================================================
FILE: README.md
================================================
# Overview
Very simply, this project demonstrates how to match an image to a bank of pre-existing images. It contains a simple front-end and image bank. The python implementation of the image bank can be easily adapted for other applications.

The image comparisons use [SURF: Speeded Up Robust Features](http://www.vision.ee.ethz.ch/~surf/eccv06.pdf) which is **scale, orientation, and to some degree affine invariant**.

A common problem in managing large numbers of images is detecting *slight* duplicates. Using a library like OpenCV which is widely available across platforms and languages is a great way to detect these duplicates.

![scale orientation invariant](http://i.imgur.com/nFASitk.gif)


# Animated description

![animation](http://i.cubeupload.com/8nVjdO.gif)


# How it works
To add an image to the bank:
- Compute SURF descriptors for the image
- Concatenate the descriptor to a "mega matrix" of pre-existing ones, making note of it's position.

To look up an image:
- Compute SURF descriptors for the image
- Perform a knn search in the "mega matrix" for the SURF descriptors found above
- For all matches, if the two are within a certain distance threshold, we increment a similary value with respect to that candidate by 1. This creates an arbitrary similarity index.
- Return the top results


The server is implemented using [flask](http://flask.pocoo.org/) and the front end uses [react](http://facebook.github.io/react/)


# Install:
## OSX
Need to install `opencv` and `imagemagick` (todo: add links)
```sh
pip install sqlite3
pip install numpy
pip install flask
pip install wand
pip install flask
npm install
```

# Development:
compile front end
`webpack`

watch for changes on front end
`webpack --watch`

run server:
`python server.py`

watch for changes on server:
uncomment this line in `server.py` `app.debug = True`
**note: this is by default on**

# Optimization:
- The implementation is poorly optimized, there is a rudimentary attempt to distribute the "mega matrix" to take advantage of multiple cores. At any sort of scale, you probably want to look into doing some sort of distributed nearest neighbor search.

- By default the server persists the bank data in `bank.db` which is a simple sqlite database with pickled python objects. This is merely for convenience between server restarts. While it is running, the server keeps everything in local memory.

# Related projects:
- [isk-daemon](https://github.com/ricardocabral/iskdaemon)

# Notes:

- Tested with around 200k images without issues.

- This is only tested on OS X Mavericks, it shouldn't have any problems on linux. It is completely untested on windows.

- [A Sample dataset](http://www.vision.caltech.edu/Image_Datasets/Caltech256/). untar it and just POST them all to the server `find <MY_DATASET_DIR> -name "*.<IMAGE_EXTENSION>" -exec curl -i -F file=@{} \;`


# LICENSE
**mineye** source code is released under the **MIT License**

The **SURF and SIFT algorithms implemented by OpenCV are patented** You will have to switch out the feature detector for something else.


================================================
FILE: img.py
================================================
import cv2
import numpy
import sqlite3
import pickle
from datetime import datetime


#max number of images in each matrix, for parallel processing
DESC_MAX_LEN = 100000
#sqlite db for persistence
BANK_FILENAME = 'bank.db'

'''
note the licensing issues with using SURF/SIFT, alternatives are FREAK, BRISK for
feature detection
'''
def get_surf_des(filename):
    f = cv2.imread(filename)
    #hessian threshold 800, 64 not 128
    surf = cv2.SURF(800, extended=False)
    kp, des = surf.detectAndCompute(f, None)
    return kp, des

def get_conn():
    return sqlite3.connect('bank.db')

class _img:
    def __init__(self):
        self.imap = []
        self.r = 0
        self.descs = []
        index_params = dict(algorithm=1,trees=4)
        self.flann = cv2.FlannBasedMatcher(index_params,dict())

    def add_image(self, filename, des=None):
        if des == None:
            kv, des = get_surf_des(filename)
        self.imap.append({
            'index_start' : self.r,
            'index_end' : self.r + des.shape[0] - 1,
            'file_name' : filename
        })
        self.r += des.shape[0]
        #it's really slow to do a vstack every time, so just maintain a list and
        #replicate it as a concatenated numpy ndarray every time. an optimization
        #would be to do a numpy.vstack((self.descs, numpy,array(des))) where self.descs
        #is a numpy.array
        self.descs.append(des)

    def match(self, filename, limit=20):
        kp, to_match = get_surf_des(filename)
        img_db = numpy.vstack(numpy.array(self.descs))
        #this should be reversed, need to update distance calculation
        matches = self.flann.knnMatch(img_db, to_match, k=4)
        sim = dict()
        for img in self.imap:
            sim[img['file_name']] = 0
        for i in xrange(0, len(matches)):
            match = matches[i]
            if match[0].distance < (.6 * match[1].distance):
                for img in self.imap:
                    if img['index_start'] <= i and img['index_end'] >= i:
                        sim[img['file_name']] += 1
        return sim

    def __len__(self):
        return len(self.descs)

class img:
    def __init__(self):
        self.ims = [_img()]
        self.count = 0

    def get_count(self):
        return self.count

    def add_image(self, filename, des=None):
        self.count += 1
        self.ims[-1].add_image(filename, des=des)
        if len(self.ims[-1]) > DESC_MAX_LEN:
            self.ims.append(_img())

    def match(self, filename, limit=20):
        import multiprocessing.dummy
        p = multiprocessing.dummy.Pool(10)

        def f(instance):
            return instance.match(filename, limit=limit)

        res = p.map(f, [i for i in self.ims])
        sim = dict((k,v) for d in res for (k,v) in d.items())
        sorted_sim = sorted(sim.items(), key=lambda x:x[1], reverse=True)[0:limit]
        sorted_sim = [{'image' : x[0], 'similarity' : x[1]} for x in sorted_sim]
        sorted_sim = filter(lambda x:x['similarity'] > 5, sorted_sim)
        return sorted_sim

class persisted_img(img):
    def __init__(self):
        #optimization, should additionally wrap img once more instead, so it works without persistence
        img.__init__(self)
        with get_conn() as conn:
            c = conn.cursor()
            c.execute('''CREATE TABLE IF NOT EXISTS descs
                        (filename, des,kp)
                        ''')
            conn.commit()
            c.execute(
                '''
                SELECT filename,des
                FROM descs
            ''')
            while True:
                row = c.fetchone()
                if not row:
                    break
                filename = row[0]
                des = pickle.loads(str(row[1]))
                print 'img.__init__: loading descriptor for file %s from db' % (filename)
                if des == None:
                    print 'img.__init__: error loading descriptor for %s from db' % (filename)
                    continue
                self.add_image(filename, des=des)

    def add_image(self, filename, des=None):
        if des == None:
            kv, des = get_surf_des(filename)
            with get_conn() as conn:
                c = conn.cursor()
                data = sqlite3.Binary(pickle.dumps(des, pickle.HIGHEST_PROTOCOL))
                c.execute('''
                    INSERT INTO descs(filename, des) VALUES (?,:data)
                    ''',
                    [filename, data]
                )
                print 'INSERT  %s to db' % (filename)
                conn.commit()
        img.add_image(self, filename, des=des)




================================================
FILE: package.json
================================================
{
  "name": "similarity",
  "version": "0.0.0",
  "description": "",
  "main": "webpack.config.js",
  "dependencies": {
    "jsx-loader": "~0.9.0",
    "css-loader": "~0.6.12",
    "less-loader": "~0.7.2",
    "less": "~1.7.0",
    "style-loader": "~0.6.3",
    "envify": "~1.2.1",
    "react": "~0.10.0",
    "superagent": "~0.17.0",
    "lodash": "~2.4.1",
    "jquery": "~2.1.0"
  },
  "devDependencies": {},
  "scripts": {
    "test": "echo \"Error: no test specified\" && exit 1"
  },
  "author": "",
  "license": "ISC"
}


================================================
FILE: pages/home.jsx
================================================
/** @jsx React.DOM */
/*NPM includes*/
var _ = require('lodash');
var $ = require('jQuery');

/*React includes*/
var React = require("react/addons");
var cx = React.addons.classSet;

/*Styling*/
require('./home.less');

var UploadForm = React.createClass({
  handleClick : function(e){
    var node = this.getDOMNode();
    onComplete = this.props.onComplete;

    e.preventDefault();
    var formData = new FormData(this.getDOMNode());
    $.ajax({
        type:$(node).attr('method'),
        url: $(node).attr('action'),
        data:formData,
        cache:false,
        contentType: false,
        processData: false,
        success:function(data){
          onComplete(data);
        },
        error: function(data){
          alert('error uploading!');
        }
    });
  },
  render : function(){
    return (
      <form method={this.props.method} action={this.props.action} encType='multipart/form-data' onSubmit={this.handleClick}>
        <input type="file" name="file" />
        <input type="submit" value={this.props.value} />
      </form>
    );
  }
});

var View = React.createClass({
  getInitialState : function(){
    return {
      bank : [],
      count : 0,
      related : []
    };
  },
  refreshBank : function(){
    $.getJSON('/bank', function(bank){
      if(!bank || !bank.latest || !bank.count){
        return;
      }
      this.setState({
        bank : bank.latest,
        count : bank.count
      });
    }.bind(this));
  },
  relatedImagesChanged : function(images){
    if(!images || !images.length){
      return;
    }
    images = JSON.parse(images);
    this.setState({
      related : images
    });
  },
  componentDidMount : function(){
    this.refreshBank();
  },
  render : function(){
    var images = _.map(this.state.bank, function(image){
      return (
        <li key={image}>
          <img src={image} className="thumb-100" />
        </li>
      );
    });
    var related = _.map(this.state.related, function(image){
      return (
        <li key={image.image}>
          <img src={image.image} className="thumb-100" />
          <h5>similarity : {image.similarity}</h5>
        </li>
      );
    });
    return (
      <div>
        <h1>Upload a file to check!</h1>
        <UploadForm action="/lookup" value="Upload" method="POST"  action="/similar" onComplete={ this.relatedImagesChanged }/>
        <h1>Results</h1>
        <p>Please upload a file to get some matches</p>
        <ul className="thumbs">
          {related}
        </ul>
        <h1>Recent Images Uploaded</h1>
        <ul className="thumbs">
          {images}
        </ul>
        <h3>Total Number of Images: {this.state.count} </h3>
        <h1>Add a file to the bank!</h1>
        <UploadForm action="/bank" value="Upload to bank" method="POST" action="/bank" onComplete={ this.refreshBank }/>
      </div>
    )
  }
});

React.renderComponent(
  View(),
  document.getElementById('content')
);


================================================
FILE: pages/home.less
================================================
.thumb-100{
  width: 100px;
  height: 100px;
}

ul.thumbs{
  >li{
    display: inline-block;
    margin-left: 1px;
  }
}


================================================
FILE: server.py
================================================
from os import listdir
from os.path import isfile, join
import traceback
import json
import uuid
import re
import tempfile
from flask import Flask, request
import wand.image
import wand.display
import wand.exceptions
app = Flask(__name__)

#local stuff
from img import persisted_img
im = persisted_img()

BANK_PATH = 'static/img/bank'
BANK_THUMB_PATH = join(BANK_PATH,'thumb')
print 'USING BANK PATH ' + BANK_PATH
print 'USING THUMB PATH ' + BANK_THUMB_PATH

def get_images(path):
    #this isn't very robust, oh well
    return filter(
        lambda x : re.search('\.(jpg|jpeg|png)', x.lower()) != None,
        [join(path, f) for f in listdir(path) if isfile(join(path,f))]
    )

def get_bank_images():
    return get_images(BANK_PATH)

def get_thumb_images():
    return get_images(BANK_THUMB_PATH)

@app.route("/")
def index():
    return '''
        <html>
            <head>
            </head>
            <body>
              <div id="content"></div>
              <script type="text/javascript" src="/static/js/all.js"></script>
            </body>
        </html>
        '''

@app.route('/similar', methods=['POST'])
def similar():
    if request.method == 'POST':
        file = request.files['file']
        if file:
            tmpfile = join(
                tempfile.gettempdir(),
                file.name
            )
            file.save(tmpfile)
            #lol shitty
            try:
                with wand.image.Image(filename=tmpfile) as img:
                    img.resize(256, 256)
                    img.save(filename=tmpfile)
                matches = im.match(tmpfile, limit=10)
                return json.dumps(matches)
            except:
                traceback.print_exc()
                pass
    return '', 400

@app.route('/bank', methods=['GET', 'POST'])
def bank():
    if request.method == 'POST':
        file = request.files['file']
        print file
        if file:
            tmpfile = join(
                tempfile.gettempdir(),
                file.name
            )
            guid = str(uuid.uuid4().get_hex().upper()[0:12]) + '.jpg'
            dstfile = join(
                BANK_PATH,
                guid
            )
            dstfile_thumb = join(
                BANK_THUMB_PATH,
                guid
            )
            file.save(tmpfile)
            try:
                with wand.image.Image(filename=tmpfile) as img:
                    img.save(filename=dstfile)
                    #will potentially produce some funny results with extremely wide/oblong images
                    img.resize(256, 256)
                    img.save(filename=dstfile_thumb)
                    im.add_image(dstfile_thumb)
            except wand.exceptions.MissingDelegateError:
                return 'input is not a valid image', 500
            return '', 200

    elif request.method == 'GET':
        limit = 10
        try:
            limit = int(request.args.get('limit', '10'))
        except ValueError:
            pass
        #note, will spit back any non dir
        files = get_bank_images()
        return json.dumps({
            'count' : im.get_count(),
            'latest' : ['/'+f for f in files[0:limit]]
            })
    return '', 400

if __name__ == "__main__":
    #todo: toggle debug from config
    app.debug = True
    app.run()


================================================
FILE: static/img/.gitignore
================================================
*
!.gitignore


================================================
FILE: static/js/.gitignore
================================================
all.js


================================================
FILE: webpack.config.js
================================================
module.exports = {
  context: __dirname,
  entry: './pages/home.jsx',
  output: {
    path: __dirname + '/static/js',
    filename: 'all.js'
  },
  module: {
    loaders: [
      {test: /\.jsx$/, loader: 'jsx-loader'},
      {test: /\.less$/, loader: 'style-loader!css-loader!less-loader'},
      {test: /bower_components\.*\.js$/, loader: "script-loader"}
    ]
  }
};