Fruit-Classification: Work-In-Progress

Fruit Classification/Identififcation using TensorFlow-Keras on Fruits 360 dataset

Understand Dataset:

Step 1 - EDA:

Method/Code Snippet:

#get path to root dir
base_dir_path = os.getcwd()

#build path to train dir
train_dir_path = os.path.join(base_dir_path,'train')

#build path to test dir
test_dir_path = os.path.join(base_dir_path,'test')

readData(base_dir_path)

Console Output:

Total Number of Classes in train DataSet:  95
Total Number of Classes in test DataSet:  95
Total Number of train samples:  48905
Total Number of test samples: 16421

Method/Code Snippet:

understandData(base_dir_path,'train')

Console Output:

CLASS NAME          NUMBER OF IMAGES
Apple Braeburn      492
Apple Golden 1      492
Apple Golden 2      492
Apple Golden 3      481
Apple Granny Smith  492
.
.
.

Build Model and Train Dataset:

Approch:

• I used MobileNetV2 architecutre, pre-trained on ImageNet dataset for feature extraction.
• Next I use these features and ran through a new classifier, which is trained from scratch.
• As stated in my Favourite Book: Deep Learning with Python. We took convolutional base(conv_base) of MobileNetV2, ran new data through it and trained a new classifier on top of the output.
• So basically, I extended the conv_base by adding Dense layer followed by DropOut layer, and running whole network on input data with data augmentation.
• Well!! this is computationally bit expensive, but meh!! I have enough processing power.
• Important Thing, I freeze the convolutional base so as to avoid updating their weights.

Step 2 - Compiling Model:

Method/Code Snippet:

#Get list of All classes
AllClassNames = getAllClassNames(train_dir_path)
num_of_classes = len(AllClassNames)

#build dict of clas_id and classname
DictOfClasses = {i : AllClassNames[i] for i in range(0, len(AllClassNames))}

#Compile classification model
classifyModel=compileClassifyModel(num_of_classes)

Console Output:

Layer (type)                 Output Shape              Param #   
=================================================================
mobilenetv2_1.00_224 (Model) (None, 1280)              2257984   
_________________________________________________________________
flatten_1 (Flatten)          (None, 1280)              0         
_________________________________________________________________
dense_2 (Dense)              (None, 512)               655872    
_________________________________________________________________
dropout_1 (Dropout)          (None, 512)               0         
_________________________________________________________________
dense_3 (Dense)              (None, 95)                48735     
=================================================================
Total params: 2,962,591
Trainable params: 704,607
Non-trainable params: 2,257,984
_________________________________________________________________

Step 3 - Training compiled Model:

Method/Code Snippet:

#Start training model on train dataset
trainingHistory,trainedModel_filename = trainClassifyModel(classifyModel)

#Plot the training results
plotTrainResults(trainingHistory)

Training Results:

Epcohs:20

Step 4 - Prediction:

Method/Code Snippet:

#path to test image
ImagePath = 'test/Banana Red/99_100.jpg'

#path to trained-saved model
path_trained_model = os.path.abspath(trainedModel_filename)

#load trained model
trainedModel = getTrainedModel(path_trained_model)

#perform predictions
AllProbs = predictFruitClass(ImagePath,trainedModel,DictOfClasses)

Console Output:

Banana

Issues and Challenges:

• Need more diverse data for each fruit class.
• It is really hard for model to infer the type of fruit, this may be due to closer properties(shape,color etc) of the object. I mean, it is easier for model to recognise Banana compared to other fruit class.
• For Example, model predicts Grape White as Guava. Refer Above grid image. This has been observed with several other fruit classes.

TODO:

• Test with more epochs.
• Test with ResNet, InveptionV3,Xception models
• Add method to print Top-K predicted categories/classes.
• Add method to Evaluate prediction accuracy and loss on whole test dataset.

Refrences:

• MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications https://arxiv.org/abs/1704.04861
• https://keras.io/applications/#resnet
• Deep Learning with Python, François Chollet.