Edit this page on GitHub

Image Annotation

Image Classification

ImageNET Concepts - Convolutional Neural Networks

A tool is available in the cluster to classify images with ImageNET concepts (a total of 1000 concepts - List available here). The current version performs classification using the state-of-the-art GoogleNet CNN, which won the ILSVRC 2014 competition with 6.65% Top-5 error.

Each image is classified within the 1000 concepts. For each concept, a probability score is assigned thus, the tool outputs the top N concepts (from the 1000) with highest probability. Namely, for each image it outputs N pairs <concept name, probability>.


Support for other CNN networks will be added later. If you find any bug or if you have any suggestion to improve the tool, please send an email to: df.semedo__at__campus.fct.unl.pt

To run the tool:

$ concepts_extraction --help
 usage: concepts_extractor.py [-h] [-n [NUM_CONCEPTS]] [-p [INPUT_PATH]]
                      [-i [SINGLE_IMG_PATH]] [-j [OUT_JSON]]
 CNN ImageNet Concept extractor.
 - Current version only supports GoogleNet (Inception) network.
 - Supports JPG and PNG images (w/ 3 channels).

 Author: David Semedo (df.semedo__at__campus.fct.unl.pt )

 optional arguments:
  -h, --help            show this help message and exit
  -n [NUM_CONCEPTS], --num_concepts [NUM_CONCEPTS]
                       Number of concepts to retrieve.
  -p [INPUT_PATH], --input_path [INPUT_PATH]
                       Path with target images.
                       Path of a single target image.
  -j [OUT_JSON], --out_json [OUT_JSON]
                       Store output in a JSON file.

To extract 20 concepts per image, from a folder and store the output in a JSON file \<out.json\>:

$ concepts_extraction -n 20 -p ` -j `