Design and Deployment of an Image Polarity Detector with Visual Attention

This repository concerns the experimental campaign of the paper "Design and Deployment of an Image Polarity Detector with Visual Attention" by Ragusa et al., DOI

Authors

• Edoardo Ragusa: [email protected]
• Tommaso Apicella: [email protected]
• Christian Gianoglio: [email protected]
• Rodolfo Zunino: [email protected]
• Paolo Gastaldo: [email protected]

Table of contents

• Python project
• Android application
• Reference

Python project

Requirements

The requirements to run the python code are the following:

• Python 3.6
• Tensorflow 1.14
• Tensorflow Slim
• Tensorflow Object Detection API
• Numpy
• OpenCV
• Pillow
• scikit-learn

Description

The pythonProject enables to run the algorithm described in the paper on a test set. The module consists of 3 files:

• box_detection.py: methods to perform object detection for a single image. These functions can return boxes images and/or boxes coordinates with a score greater than threshold (in some cases set to 0.3) .
• write_file.py: methods to clear a file, write a text, write accuracy or confusion matrix into a text file.
• polarity_detection_visual_attention.py: this script employs the algorithm described in the paper to detect image polarity using visual attention.

Model checkpoints are available:

• mobilenet_v1_1.0_224_anp40: entire image model.
• mobilenet_v1_1.0_224_anp40_patches: patches model.
• ssd_mobilenet_v1_ILSVRC: saliency detector ILSVRC.
• ssd_mobilenet_v1_SOS: saliency detector SOS.

Android application

Requirements

Currently, the app settings in build.gradle file are:

• compileSdkVersion 29
• minSdkVersion 19
• targetSdkVersion 29

The minimum SDK 19 indicates that smartphones with an Android version higher (or equal) than Android 4.4 (KitKat) can run the application.

Description

This demo enables to run the algorithm described in the paper on Android smartphone.
User can select an image from the asset or from storage.
The app can run inference using only the polarity classifier or using the saliency detector in combination with polarity classifier.
For each performed inference phase the app displays:

• Inference time: time spent to perform forward phase.
• Inference result: eventual polarity output, "Positive" or "Negative".
• Inference confidence: confidence associated with the prediction. A number in [0, 1] range if the final polarity class is established by the entire image classifier, "---" otherwise. The rationale behind this is that the algorithm's operations can change the range and the meaning of the confidence value. For more details, please refer to the paper.

Each model of the pythonProject was converted to TFLite exploiting post-training quantization process for FP32, FP16 and INT8 formats. Models are available in the asset folder.
The most important directories are:

• assets: collection of test image, saliency and polarity models, labels file.
• utils: holds java classes of algorithm, file and general utilities.
• activity: holds MainActivity, which is the only activity of the application.
• classes: holds java classes concerning the prediction, the polarity classifier and saliency detector.

Reference

If you find the code or pre-trained models useful, please cite the following paper:

Design and Deployment of an Image Polarity Detector with Visual Attention. E. Ragusa, T. Apicella, C. Gianoglio, R. Zunino and P. Gastaldo. Cognitive Computation, 2021. DOI

@article{ragusa2021design,
  title={Design and Deployment of an Image Polarity Detector with Visual Attention},
  author={Ragusa, Edoardo and Apicella, Tommaso and Gianoglio, Christian and Zunino, Rodolfo and Gastaldo, Paolo},
  journal={Cognitive Computation},
  pages={1--13},
  year={2021},
  publisher={Springer}
}