The models used for detection are SSD [SSD: Single Shot MultiBox Detector](https://arxiv.org/abs/1512.02325) and YOLOv3 [YOLOv3: An Incremental Improvement] (https://arxiv.org/abs/1804.02767), they are imported from the following repositories:
Grab the pretrained weights of SSD and YOLO3 from [Drive_Weights](https://drive.google.com/drive/folders/1LSc9FkAwJrAAT8pAUWz8aax_biFAMMXS?usp=sharing)
The images used for the design of this model were extracted by air analysis, specifically: FLIR aerial radiometric thermal infrared pictures, taken by UAV (R-JPEG format). Which were converted into .jpg images for the training of these detection models.
By the end of this process, the code will write the weights of the best model to file best_weights.h5 (or whatever name specified in the setting "saved_weights_name" in the config.json file). The training process stops when the loss on the validation set is not improved in 20 consecutive epoches.
### 4. Perform detection using trained weights on image, set of images
It carries out detection on the image and write the image with detected bounding boxes to the same folder.
## Evaluation
The evaluation is integrated into the training process, if you want to do the independent evaluation you must go to the folder ssd_keras-master or keras-yolo3-master and use the following code
`python evaluate.py -c config.json`
Compute the mAP performance of the model defined in `saved_weights_name` on the validation dataset defined in `valid_image_folder` and `valid_annot_folder`.
On folder [Result yolo3 fault 1](Result_yolo3_fault_1/) show [history train](Result_yolo3_fault_1/yolo3_full_yolo.output), weight and result of this model (mAP 73.02%).
On folder [Result yolo3 fault 2](Result_yolo3_fault_2/) show [history train](Result_yolo3_fault_2/yolo3_full_yolo.output), weight and result of this model (mAP 71.93%).
On folder [Result yolo3 fault 4](Result_yolo3_fault_4/) show [history train](Result_yolo3_fault_4/yolo3_full_yolo.output), weight and result of this model (mAP 66.22%).
The idea to detect the disconnection is by calculating the luminosity of each panel, to then normalize this data and highlight the panels with a luminosity out of normality.
- Check if my changes are consistent with the [guidelines](https://github.com/RentadroneCL/model-definition/blob/master/CONTRIBUTING.md#general-guidelines-and-philosophy-for-contribution).
- Changes are consistent with the [Coding Style](https://github.com/RentadroneCL/model-definition/blob/master/CONTRIBUTING.md#c-coding-style).
- Run [Unit Tests](https://github.com/RentadroneCL/model-definition/CONTRIBUTING.md#running-unit-tests).
In ['Example_Prediction'](Code_Example/Example_prediction.ipynb) this is the example of how to implement an already trained model, it can be modified to change the model you have to use and the image in which you want to detect faults.
In ['Example_Prediction_AllInOne'](Code_Example/Example Detection AllInOne.ipynb) this is the example of how implement all trained model, you can use this code for predict a folder of images and have a output image with detection boxes.
