Skip to content

jianwu21/weed-recognization

Repository files navigation

Weed Recognization

This is the free topic for 2017-2018 block 1. You can see all experiments run by me in a jupyter notebook file.


Database description

The database include two kinds of crops,barley and wheat.For each group,we got theimages from 5 attitutes,10m,20m,30,50m.

For each image, we divide it to several small pictures(100, 100).Then we identify these pictures as weed one or crop one by using classical image processing method. Then, we use the tiny pictures as training data.

Before start

Data Generation

All imges are in .JPG format. It will take sometime for python to load it. There is one script for data generation. All image will be stored in one .db file with both of RGB and HSV space. Then it will be convinient and fast for us to extract the matrix of images, which can be helpful to improve experiment speed.

python generate_data.py

python for Linux

There are some problems with python in Linux environment. So please remember to set virtual environment or virtual machine in Linux.I use virtualenv

virtualenv env
source env/bin/activate

For preprocess our images, I choose OpenCV

CNN is so hot now.It made something easier.Instead of old school image processing method,cnn seems a better for us to do area recognition.

For CNN model building,. I chose Keras.

Preprocess for images

  • Color normalization
  • Images normalization
  • Remove the edge with value 0

Color normalization

Since our pictures are taken from different weather and different angle. So the light may have significant influence in the images for training and test. Converting it from RGB space to HSV space is a good idea for the origin images to be normalized. Because HSV/HSL(Hue, Saturation, Value/Lightness) space can express the relationship with light directly.

Here, the effect is concentrated both in the saturation component and the value component, whereas the hue component is hardly affected.We choose to correct both saturation S and value V using a linear fit to each image.

Image normalization

Make all images including traing set and testing set be in the same scale

Data Training

  • Training
  • Validation
  • Test

Procedure for Training and Validation

Since the file in the original folder include labeled images and surround images. But the training for us is just the labeled RGB images(100*100). The idea for that is to store label, RGB channel and HSV channel in one sqlite file. Then you can extract them for training anytime. That you can do that by,

python generate_data.py

After doing that, now we can start training. The classical method for model selection is that we train one classifier based on each randomly picked training images group. Then we pick the best one. Now, since the limit for the time and hardware, I just train one classifier. I will update the amount of training group based on one improved CNN after talking with Francois.

Test

For the test, we want to use classifier in one entire picture. Firstly, the size of one original images is 3000 4000. So we should divide the testing original picture to 30*40 batches, every batch's size is 100*100. Then label them depending on selected best model.

CNN structure

Run python CNN_mmodel.py to build CNN model. The specific info is displayed below.

Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_1 (Conv2D)            (None, 100, 100, 64)      1792      
_________________________________________________________________
activation_1 (Activation)    (None, 100, 100, 64)      0         
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 50, 50, 64)        0         
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 50, 50, 128)       73856     
_________________________________________________________________
activation_2 (Activation)    (None, 50, 50, 128)       0         
_________________________________________________________________
conv2d_3 (Conv2D)            (None, 50, 50, 128)       147584    
_________________________________________________________________
activation_3 (Activation)    (None, 50, 50, 128)       0         
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 25, 25, 128)       0         
_________________________________________________________________
dropout_1 (Dropout)          (None, 25, 25, 128)       0         
_________________________________________________________________
flatten_1 (Flatten)          (None, 80000)             0         
_________________________________________________________________
dense_1 (Dense)              (None, 10)                800010    
_________________________________________________________________
activation_4 (Activation)    (None, 10)                0         
_________________________________________________________________
dropout_2 (Dropout)          (None, 10)                0         
_________________________________________________________________
dense_2 (Dense)              (None, 2)                 22        
_________________________________________________________________
activation_5 (Activation)    (None, 2)                 0         
=================================================================
Total params: 1,023,264
Trainable params: 1,023,264

Logs of training

I use tensorfolw to record the logs during model training. So you can also check that by using

tensorboard --logdir=./logs

model review

You can check how the model work in notebook

Final Report

You can read the final report for more details. You can also download it if you can not read it online.

About

No description or website provided.

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages