From vast cities to dense forests, fire incidents pose a major threat to the world. The information derived from global statistics for various regions of the world regarding losses resulting from fires indicates that the current technology used to detect fires has failed to eliminate the deaths, injuries, and economic losses resulting from fires. The primary focus of this paper is on addressing challenges such as discerning fires from non-fire elements, adapting to diverse lighting conditions, and achieving timely detection. This paper aims to enhance the accuracy of fire detection and localization in digital images through deep learning techniques. We propose the utilization of a modified ResNet50 network specifically tailored to localize fires rather than merely detecting them. To achieve this, input images are partitioned into non-overlapping blocks of 40 × 40 pixels. Each block is treated as an individual image and classified as either a fire or non-fire block. All the output classified blocks are inserted in a new image with the same location as the original image, with changing in the color of non-fire blocks to black color. This process helps the ResNet50 network to work as classifies the image and determines the location of the interested object. The dataset used in the current work is selected from Kaggle. The proposed model exhibits exceptional performance, with a training set accuracy of 99.77% and a test accuracy of 99.12%. Comparative analysis with other similar works demonstrates superior results, highlighting the efficacy of our approach.