Background: The forest environment is intricate and dynamic, with its depiction influenced by various factors such as geographical location, weather conditions, and capture angles. Relying solely on flame or smoke is insufficient for precise fire information. Aims: This paper proposes a method for accurate detection on forest flame and smoke based on improved feature extraction module with enhanced image processing. Methods: We a fusion-guided filtering image processing method and flame segmentation strategy to augment the quality of dataset. Additionally, an outstanding extraction backbone, incorporating ghost modules and decoupled fully connected (DFC) attention modules, is developed to increase the model’s receptive field. Furthermore, the ELAN-S neck with SimAM attention mechanism is introduced to fuse features from the backbone network, facilitating the extraction of shallow and deep-level semantic information. Key results: Compared to YOLOV7, our model demonstrates superior performance with a 5% increase in mean average precision (mAP), a 4.3% increase in average precision for small objects (APS), and a 3–4% enhancement in other metrics. Conclusions: The proposed model achieves a good balance between detection speed and detection accuracy. The improved model performs well in real forest fire detection scenarios. Implications: In the early forest fire detection, the model considers both flame and smoke information to describe the fire situation, and effectively combines the semantic information of both for fire warning.

UAVs are commonly used in forest fire detection, but the captured fire images often suffer from blurring due to the rapid motion between the airborne camera and the fire target. In this study, a multi-input, multi-output U-Net architecture that combines spatial domain and frequency domain information is proposed for image deblurring. The architecture includes a multi-branch dilated convolution attention residual module in the encoder to enhance receptive fields and address local features and texture detail limitations. A feature-fusion module integrating spatial frequency domains is also included in the skip connection structure to reduce feature loss and enhance deblurring performance. Additionally, a multi-channel convolution attention residual module in the decoders improves the reconstruction of local and contextual information. A weighted loss function is utilized to enhance network stability and generalization. Experimental results demonstrate that the proposed model outperforms popular models in terms of subjective perception and quantitative evaluation, achieving a PSNR of 32.26 dB, SSIM of 0.955, LGF of 10.93, and SMD of 34.31 on the self-built forest fire datasets and reaching 86% of the optimal PSNR and 87% of the optimal SSIM. In experiments without reference images, the model performs well in terms of LGF and SMD. The results obtained by this model are superior to the currently popular SRN and MPRNet models.