Ecosyste.ms: Awesome

An open API service indexing awesome lists of open source software.

Awesome Lists | Featured Topics | Projects

https://github.com/dlsaavedra/detector_gdxray

This repository contains 4 detectors for use in the detection of dangerous elements in X-.ray images
https://github.com/dlsaavedra/detector_gdxray

Last synced: 2 months ago
JSON representation

This repository contains 4 detectors for use in the detection of dangerous elements in X-.ray images

Awesome Lists containing this project

README 

        
# Detector_GDXray

This repository contains 4 detectors for use in the detection of dangerous elements in X-ray images.

The repositories of the detectors used are the following:



  + [Yolov2](https://github.com/experiencor/keras-yolo2)

  + [Yolov3](https://github.com/experiencor/keras-yolo3)

  + [SSD300 y SSD7](https://github.com/pierluigiferrari/ssd_keras)

  + [RetinaNet](https://github.com/fizyr/keras-retinanet)



To start using the detectors you just have to clone or download the repository and follow the instructions in the Readme.



# Requirement

To install the dependencies, run



`pip3 install -r requirements.txt`



And if you want use the GPU to work, make sure you've got the drivers installed beforehand (CUDA).

and install

`pip3 install tensorflow-gpu tensorflow-gpu==1.13.1`



It has been tested to work with Python 3.6.8.



# 0- Data GDXray

The database to use the public database GDXray which we can download from [GDXray](https://domingomery.ing.puc.cl/material/gdxray/), specifically you must create the "Data" folder and copy the folders B0049, B0051, B0052, B0076, B0082 and BX_100 (new folder) from GDXray. For the testing use the folder B0046 with its annotation, you can also download these folders in the following link :

[Data Link](https://drive.google.com/drive/folders/1edU-QUHdrRY6-xgmeL0WIvb8PiH9xRHz?usp=sharing).



Example image test:



![Test](images/B0046_0004.png)



# 1- Use Trained Model

## Download the pre trained weights

 First you must download the pre-trained weights in Imagenet given in the following link [Pre-Trained Weights](https://drive.google.com/drive/u/0/folders/1ffxuT9TBz9ZZDgRM0uzZ7Ffi4CWEwdNf) and place them in their respective folders.



### Download Weights Trained Model

You can use already trained models to use directly in X-ray images to detect "Gun, Knife, Razor and Shuriken", for this create the "Weights Trained Model" folder in the repository folder and download the weights in the folder [Weights Trained Model](https://drive.google.com/drive/u/0/folders/1B832BTuheU7pNzSxMFWU_f5Xkwu70bn2).

Follow the following commands for detection:



+ `python predict_yolo2.py -c config_full_yolo2_infer.json -i input_path/folder -o save/folder/detection` Yolov2

+ `python predict_yolo3.py -c config_full_yolo3_infer.json -i input_path/folder -o save/folder/detection` Yolov3

+ `python predict_ssd.py -c config_7_infer.json -i input_path/folder -o save/folder/detection` SSD7

+ `python predict_ssd.py -c config_300_infer.json -i input_path/folder -o save/folder/detection` SSD300

+ `python predict_retinanet.py -c config_resnet50_infer.json -i input_path/folder -o save/folder/detection` RetinaNet



Example Detection:



+ `python predict_yolo2.py -c config_full_yolo2_infer.json -i Data/Testing/images/B0046_0001.png -o Example_yolo2/`

+ `python predict_retinanet.py -c config_resnet50_infer.json -i Data/Testing/images/B0046_0001.png -o Example_retinanet/`



![Train](images/Detection.png)



# 2- Create own Data set for Training



To create a database to train detection models, you must run the command `python Create_Data_Superimpose.py`, if you want to modify the amount of data you must modify the parameters of the file `Create_Data_Superimpose.py`



You can also download a instance experiment 0 in the following link :[Experiment_0](https://drive.google.com/drive/folders/1aY2bMXvZ7qYl16NE7bo0yLseahbfJAql?usp=sharing).



Example image created:



![Train](images/Bolso4_5_100_5_0.png)



# 3- Training



 ## Run Training

 Following commands for Trainig:



+ `python train_yolo2.py -c config_full_yolo2.json` Yolov2

+ `python train_yolo3.py -c config_full_yolo3.json` Yolov3

+ `python train_ssd.py -c config_7.json` SSD7

+ `python train_ssd.py -c config_300.json` SSD300

+ `python train_retinanet.py -c config_resnet50.json` RetinaNet



To change the folders used for training and the parameters of this, you must modify the configuration file `.json`.



 ## Suggestion

 If your computer does not have a GPU, it is recommended to only train the SSD7 model, due to the computation time that the other models need. Also, need to change configuration file of YOLO3`config_full_yolo3.json` line `gpu: "0,1"` to `gpu: "0"` and for Retinanet `"multi-gpu" :  1`.



# 4- Evaluate

For evaluate the trained models just need specify the test folder with respective annotation in the configuration file `json`.



 ## Run Evaluate

 Following commands for Evaluate:

 + `python evaluate_yolo2.py -c config_full_yolo2.json` Yolov2

 + `python evaluate_yolo3.py -c config_full_yolo3.json` Yolov3

 + `python evaluate_ssd.py -c config_7_infer.json` SSD7

 + `python evaluate_ssd.py -c config_300.json` SSD300

 + `python evaluate_retinanet.py -c config_resnet50.json` RetinaNet



The default configuration have folder B0046 as folder test.



# 5- Usage your Models

For use your trained models to detect, follow the same structure that in section 1.



+ `python predict_yolo2.py -c config_full_yolo2.json -i input_path/folder -o save/folder/detection` Yolov2

+ `python predict_yolo3.py -c config_full_yolo3.json -i input_path/folder -o save/folder/detection` Yolov3

+ `python predict_ssd.py -c config_7.json -i input_path/folder -o save/folder/detection` SSD7

+ `python predict_ssd.py -c config_300.json -i input_path/folder -o save/folder/detection` SSD300

+ `python predict_retinanet.py -c config_resnet50.json -i input_path/folder -o save/folder/detection` RetinaNet



# Contributions

Show a way to use the limited GDXray database to create a database suitable for training the presented detectors. Also simplify the use of different detectors based on the repositories mentioned.



# Research dangerous objects detection (Daniel Saavedra)

the investigation can be replicated using the same code as Experiment 0, only line 101 must be changed in `Create_Data_Superimpose.py`:

`num_image_per_baggage = 200`.

Then the training and testing of the different models is performed as explained above.