Assess Image Quality

Description

A good photo is one where factors such as backlit, blur, contrast, and composition are well balanced, creating an aesthetically pleasing image. Specifically, a good photo should have the following characteristics:

• Backlight: adjust the lighting appropriately to avoid shadows or glare on the subject in the photo. (0-100)
• Blur: the photo should be taken clearly, without blur or blur. (0-100)
• Contrast: increase the contrast in the image so that details are clearly distinguished, helping to increase bold, sharp dynamics. (0-100)
• Layout: arrange the subjects in the frame so that it is harmonious, beautiful and creates a sense of sublimation for the viewer.
  • Onethird
  • Center
  • Symmetry

The application will evaluate the image quality based on 4 criteria of backlit, contrast, blur and composition.

Installation

Window

• Clone project

    git clone https://github.com/bdts1547/AssessImage.git
  

• Download, extract model_symmetry and put it in the folder AssessImage.

  AssessImage
  |-- caffe
  |-- model
  |-- MODELS (here)
  |-- ...
  

• Create environment python 3.7.13 with Anaconda and install requirements

    conda create -n my_env python=3.7.13 -y
    conda activate my_env
    pip install -r requirements.txt
  

• Create environment python 2.7.18 with name "py27" to create sub shell for calling symmetry detection

    conda create -n py27 python=2.7.18 -y
    conda activate py27
    pip install matplotlib numpy scikit-image scipy protobuf
  

Run

• Run with streamlit

    conda activate my_env
    streamlit run streamlit_app.py
  

Demo

Input	Output

Architecture

• Backlit: feature extraction from YC_BC_R color space, then feed into SVM for backlit image quantification.
• Blur: use edge operators (Sobel, Robert, Laplacian) for feature extraction, then feed into SVM for blur image quantification.
• Contrast: an image whose pixel intensity distribution occupies less than 35% of domain [0, 255] will be considered as low contrast image.
• Layout:
  • Onethird, Center: use TRACER (RGB Salient Object Detection model) to detect objects in the image, then determine the position of objects to find the appropriate composition.
  • Symmetry: use a method for measuring symmetry in images by using filter responses from Convolutional Neural Networks (CNNs).
• Score: use the scores of backlit, blur, contrast and composition as features, then feed into the Linear Regression model for scoring.