Fire is a devastating natural disaster that affects both people and the environment. Recent research has suggested that computer vision could be used to construct a cost-effective automatic fire detection system. This paper describes a unique framework for utilizing CNN to detect fire. Convolution Neural Networks have yielded state-of-art performance in image classification and other computer vision tasks. Their use in fire detection systems will significantly enhance detection accuracy, resulting in fewer fire disasters and less ecological and social consequences. The deployment of CNN-based fire detection in everyday surveillance networks, however, is a severe problem due to the huge memory and processing needs for inference.  In this study, we offer an innovative, energy-efficient, and computationally effective CNN model for detection of fire, localization, and understanding of the fire scenario, based on the SqueezeNet architecture. It makes use of small convolutional kernels and avoids thick, fully connected layers that reduce the computational load. This paper shows how the unique qualities of the problem at hand, as well as a wide range of fire data, can be combined to make a balance of fire detection effectiveness and precision.