Weather Image Recognition Considering Light Condition Via CNNs

Project Summary

In this project, a image classification system based on convolutional neural network (CNN) is applied, which can identify five weather types: sunny, cloudy, rainy, foggy and snowy, and judge the ambient brightness (Bright & Dark) of all kinds of weather.Then, we used a dataset consisting of 8,890 weather images, including 7,899 training images and 991 test images.

CNNs We have used

BPNN

AlexNet

GoogLeNet+Inception v3

SENet+Inception v3

Weather dataset

The weather Dataset is available in this Link https://pan.baidu.com/s/1d9N89qkMuqMaydcXpoALog called 'WeatheDataset' pls contact me with email for more info || Jasonmils@whut.edu.cn

Experiment & Result

1. Data Augment

First， the data is resized and fliped for data augment

for i in [-1, 0, 1]:
    file_new = cv2.flip(file, i)
    im = cv2.resize(file, (width, height))
    train_images.append(im.reshape(1, width, height, 3) / 255.0)
    train_labels.append(label)

2. Training & Testing

Resuls are as follow.