Fruit-and-vegetables-classification

Image classification using ResNet50 model and Flask web application. 2D TSNE visualization.

To see the video about testing web-app click here: ResNet-50 web-app for image classification
The dataset you can find here.

Requirements:

cuda_11.2
python version- 3.6.2

Algorithm:

Data augmentation

I'm using tf.keras.preprocessing.image.ImageDataGenerator to make augmentation of images(spect ratio resizing, shifting, blurring, flipping)

Training a model

I'm using tf.keras.applications.resnet50.ResNet50 with input shape of the image (224, 224, 3). Weights from imagenet dataset and max pooling in layers. Before fine-tuning accuracy was 0.90023.

Fine-tuning model

As a fine-tuning I add this layers:

And the last five layers of the base model were also unfrozen and trained. After fine-tuning accuracy increased to: 0.92514.

2D visualization using TSNE

I use TSNE to depict the distribution of classes.
Result:

Web application using Flask

You can find video of testing my web application here. You can use my weights, which you can find in directory saved_model or you can train your model and save weights by running file Fruit-and-vegetables-classification