Improved YOLOv7 algorithm for flame detection in complex urban environments

Abstract To address the problems of flame detection, such as difficulties in detecting flames and poor performance in complex urban environments, an improved YOLOv7-based flame detection algorithm for complex urban scenarios is proposed. The proposed algorithm increases multi-scale feature fusion and introduces a 160 × 160 detection scale, which improves the detection capability of small target flames. Additionally, the 3 × 3 convolutions in the backbone feature extraction module of YOLOv7 are replaced with deformable convolutions (Deformable Convolution Networks v2, DCNv2), which better accommodate varying input feature map shapes and enhance the network’s learning ability in complex scenarios. Furthermore, the Convolutional Block Attention Module (CBAM) is embedded in the feature extraction module to strengthen the response to relevant features, further improving the algorithm’s performance in dynamic environments. The K-means++ algorithm is used to re-cluster the anchor boxes, enhancing the algorithm’s ability to predict target sizes and locations. The modified YOLOv7 algorithm achieves a mean Average Precision ([email protected]) of 97.1%, an improvement of 4.9 percentage points. Experimental results demonstrate that the improved YOLOv7 algorithm significantly enhances flame detection in complex urban scenarios.