Anomaly detected using aerial imagery can be of many types like unknown ship in docks, or a unidentified aircraft at the airport and other similar anomalies. In this project we've mainly focused on ship dataset to detect all the ships in the waterbody using keras-retinanet.

This code is inspired by jasperb's helping tutorial on retinanet.

firstly you should have installed nvidia cuda toolkit so your jupyter notebook can access your gpu resources.
to have your jupter notebook run on gpu, you can follow this link: https://www.techentice.com/how-to-make-jupyter-notebook-to-run-on-gpu/
you can also have a linux machine using gpu. any will work.

• os
• numpy
• tensorflow
• sklearn
• cv2
• PIL
• h5py

Fizyr has fortunately implemented the keras retinanet model and made it available to everyone who wants to use it so we don't have to implement the model from the scratch. we can fine tune the model so it works accordingly!

os.chdir(r"D:\notebooks\RetinanetTutorial")
!git clone https://github.com/fizyr/keras-retinanet.git`

firstly we need to seperate our dataset into train,test and validation.

1. strip the images of their extensions and store the names in an array
   

os.chdir(r"D:\notebooks\RetinanetTutorial/SAR Dataset/images")
data=np.empty(len(os.listdir()),dtype='object')
for i,j in zip(os.listdir(),range(len(os.listdir()))):
    data[j]="{}".format(i.strip('.png'))
print(data)

2. use sklearn to split this data in train,val and test:

train,test = train_test_split(data,test_size=0.1,shuffle=True,random_state=8)
train,val = train_test_split(train,test_size=0.2,shuffle=True,random_state=8)
print('train:{}, val:{}, test:{}'.format(len(train),len(val),len(test)))

3. store these in train.txt,val.txt and test.txt respectively:

os.chdir(r"D:\notebooks\RetinanetTutorial/SAR Dataset/ImageSets/Main")
trainf = open("train.txt","w+")
valf = open("val.txt","w+")
testf = open("test.txt","w+")
for i in train:
    trainf.write("{}\n".format(i))
for i in val:
    valf.write("{}\n".format(i))
for i in test:
    testf.write("{}\n".format(i))
trainf.close()
valf.close()
testf.close()

4. the original dataset was in the form of png. for our model we have to convert it into jpeg format. we use PIL library to convert all the images into jpeg:

os.chdir(r"D:\notebooks\RetinanetTutorial\SAR_Dataset\images")
for i in os.listdir():
    img_png = Image.open("D:/notebooks/RetinanetTutorial/SAR_Dataset/images/{}.png".format(i.strip('.png')))
    img_png = img_png.convert('RGB')
    img_png.save("D:/notebooks/RetinanetTutorial/SAR_Dataset/JPEGImages/{}.jpg".format(i.strip('.png')))

5. !mkdir D:\notebooks\RetinanetTutorial\Output\Snapshots to make a directory to store your model snapshots

os.chdir(r'D:\notebooks\RetinanetTutorial/keras-retinanet')
!pip install .

os.chdir(r'D:\notebooks\RetinanetTutorial/keras-retinanet')
!python setup.py build_ext --inplace

os.chdir(r'D:\notebooks\RetinanetTutorial/keras-retinanet')
%run keras_retinanet\bin\train.py --tensorboard-dir D:\notebooks\RetinanetTutorial\Output --snapshot-path D:\notebooks\RetinanetTutorial\Output\Snapshots --random-transform --steps 446 pascal D:\notebooks\RetinanetTutorial\SAR_Dataset

we run the train.py file which is the driver code responsible for creating our model snapshot.
-tenosrboar-dir takes the location of the output folder. this folder will have the mAP of the model after every epoch
-snapshot-path takes the locatoin of the folder in which the location of the snapshot will be stored.
--random-transform for enabling random transformation of the images for better learning
-steps provide the no. of image to take in every epoch
pascal is the type of dataset we're using. In our dataset we have dataset in the form of pascal-voc. this takes the location of the folder in which you've kept your dataset information(images,annotations,train,validation etc)

os.chdir(r"D:/notebooks/RetinanetTutorial/keras-retinanet")
%run keras_retinanet\bin\convert_model.py D:\notebooks\RetinanetTutorial\RetinanetModels\trained_model.h5 D:\notebooks\RetinanetTutorial\RetinanetModels\Inference_model.h5

we can check the map and choose the model with best mAP. you can stop the epochs when the model starts overfitting the data.

in the fizyr keras-retinanet, there is a python driver code 'testDetector.py' which we modify to take our model and our query image to detect the anamoly, which in our case is a ship

model_path = 'D:/notebooks/RetinanetTutorial/RetinanetModels/Inference_model.h5'
image_path = 'D:/notebooks/s1.jpg'
image_output_path = 'D:/notebooks/RetinanetTutorial/SAR_Dataset/detected_images/s2_detected.jpg'
confidence_cutoff = 0.5 #Detections below this confidence will be ignored

now we run this testDetector to get resultant image

you can refer to the whole driver code which follow all the sequence shown above for a better understanding and easier working.

you can use tensorboard to check on mAP values of your model after every epoch tensorboard --logdir D:\notebooks\RetinanetTutorial\Output --bind_all