Undoubtedly, the discovery of fire has made human life simpler, more enjoyable, and easier. However, the terrible damage it causes to human safety, health, and property cannot be ignored. Unpleasant fire incidents have increasingly led to significant social, economic, and human losses in recent years. The best way to prevent these incidents is to identify them in the early stages. Leading-edge (state-of-the-art) models such as Xception, Inception v3, VGG16, and VGG19 have set standards in the field of image classification. In this study, we have suggested an improved fire detection technique in images which is derived from deep convolutional neural network (ConvNets) architecture. The proposed architecture utilizes multipath convolutional layers, implying that multiple convolutional operations are performed in parallel on the same input with varying filter sizes ( \(1\times 1\) , \(1\times 3\) , \(3\times 1\) , \(2\times 2\) , \(1\times 5\) , \(5\times 1\) ) or configurations to capture a richer and more diverse feature, improving the network’s ability to understand and classify complex inputs. The results from the different paths are combined to enhance the overall feature representation. A custom dataset was employed for training, validating and testing the suggested architecture, which subsequently outperformed leading-edge models, achieving an AUC score of 98 with \(94\%\) accuracy.