This is the source code for predicting whether a image contains a cactus or not. This project is inspired from the kaggle competition which aims to build a system for autonomous surveillance of protected areas. It tasked us with creation of an algorithm that can identify a specific type of cactus in aerial imagery.

This dataset contains a large number of 32 x 32 thumbnail images containing aerial photos of a columnar cactus (Neobuxbaumia tetetzo). Kaggle has resized the images from the original dataset to make them uniform in size. The file name of an image corresponds to its id.

In this project we want to recognise whether the image contains cactus or not. For this we are going to construct a neural network with layers and train its weights. Further we are going to use Transfer Learning which speeds up training by using pre-trained classification models. We are going to train only the top layer form the pretrained layers. We are using pre-trained MoblieNetV2 model feature detector which is released Google. For training the untrained layers we use TensorFlow 2.0 optimisers.

After 30 epochs the model's validation accuracy increases form around 0.8 to 0.97. Based on the accuracy and loss graphs, more epochs may result in even greater improvements.

1. This project is inspired from Transfer Learning using Pretrained ConvNets on TensorFlow.org
2. M. Sandler, A. Howard, M. Zhu, A. Zhmonginov, L. C. Chen, MobileNetV2: Inverted Residuals and Linear Bottlenecks (2019), Google Inc.

1. Created a api which takes image as input and returns whether the image is a cacuts image or not
2. Used Flask for creating the api and heroku for server
3. The server uses pickle file of the model we created using transfer Learing.

The accuracy of the model can be further improved by finetuning the trained layers.