Discernment of Smoke in Image Frames

Smoke is a challenging object to detect because of its changing texture, color and shape etc. These features can be extracted with the help of learning algorithms like regression, SVM, decision tree, but they do not provide an optimum result when provided with the large dataset and comparatively, accuracy of the deep learning algorithm increases. The reason of the increase in the accuracy of the algorithm is that, it is trained on the provided dataset and with the increase in the number of input data, the extraction of the dominant features of the desired object will also increase, so that it can able define as well as the detect the similar object. For the detection of the smoke, convolutional neural network is used, which take an image as an input. Transfer learning is used in the algorithm, in which VGG-19 network is used but it does not provide the satisfying results, so it is modified by introducing batch normalization layers in the network. The batch normalization increases the converges rate of the network. The accuracy increases by 3% when the number of epochs increase from 10 to 50.