tcyuan373 / bcs_body_condition_score_cattle_prediction

body-condition-score_cattle prediction.

BCS_Body_Condition_Score_Cattle_Prediction

BCS Project Description:

1.Problem Statement

• Here we need to predict the BCS classes. Basically it’s a Multiclass Classification Problem.

• We have 1,2,3,4 classes.

2.Preprocessing Steps & Feature Engineering

• Here we have train , test data & labels.csv file. So based on this csv file i created a separate folders . so that each folder indicates different classes (Class_1,Class_2,Class_3,Class_4) .

• These data divided into train-valid data(each folder contains Class_1,Class_2,Class_3,Class_4 folders). You can find these in this file (Step_1_Train-Validation Split_final.ipynb).

• here we have less amount of data. We applied a data agumentation steps & also applied transformations, filters on images like Gradient, Negative Images, Box filter , Adaptive + Gassiun thresholding , Discrete Fourier Transform , Est. transformation , Fitting Polygons , log_transformation , Gassiuan filters + Kernals , Canny_edge , Image_temparature , Contrast Stretching ,K-means , Keras augmentation.

• You can find all these image preprocessing steps in (Image Preprocessing.ipynb ) file.

• After applying all the above feature engineering steps finally we’re with 2466 images of train,719 images of valid.

• Data Visualization analysis is available in (visualization.ipynb) file

3. Models Applied & Results

• We used a base CNN model along with Reguralization techniques(Drop out , Batch Normalization , MaxPooling) etc.

• We have used activation functions ( Relu, swish) . and also used softmax as in the last layer.

• Base CNN model given Train_Accuracy of 73.01% , Validation_Accuracy of 69.96% (After applying all filters & Keras augmentation).

• You Can see this model in (Base_CNN.ipynb) file

• After this we tried with transfer learning techniques VGG16 , Resnet101.

• The same preprocessed images are given input to these above transfer techniques.

• VGG16 has given Train_accuracy of 72.43% , Valid_accuracy of 68.98% (After applying all filters & Keras augmentation).

• You Can see this in (VGG16.ipynb) file.

• Resnet has given Train_accuracy of 70.85% , valid_accuracy 69.40% (After applying all filters & Keras augmentation).

• You can see this in (Resnet101.ipynb) file.

• You can see some of the model analysis results images are store in Output_images folder.

Base Data Available in this link : https://drive.google.com/file/d/1SJYEUO1IukwxP42BAi8Cr0n_QX5G1sT4/view?usp=drive_web

Models files & Preprocessed files are empty, not loaded to github due to large file size.We can create those files by running the notebooks , process mentioned above.

Sample Data View.

Some of the research documnets links :

1.https://www.uaex.edu/publications/pdf/fsa-4008.pdf

2.https://www.grandin.com/dairy.cow.photo.charts.html

3.https://www.vet.cornell.edu/sites/default/files/1e_Elanco%20Cow%20Body_condition_scoring_V3.pdf