Abstract: Forest fires can spread rapidly to threaten forest ecosystems, nearby communities, and emergency-response resources. Unmanned aerial vehicles (UAVs) can offer flexible observation for large and inaccessible forest areas; However, fire-spot detection is still challenging, particularly when thick smoke weakens the clarity of visible textures and non-fire hot objects interfere with infrared images. In this study, the dual-modal fusion focus YOLO (DFF-YOLO) framework was proposed to detect smoke-interfered fire spots from visible-light and infrared images, based on the YOLOv11n baseline. Two parallel backbone branches were used to preserve complementary temperature and texture information. Channel prior convolutional attention (CPCA) was placed inside each branch to purify and enhance modality-specific features. Multi-scale depthwise separable convolution and channel-spatial attention were integrated beyond direct fusion. Cross-modal interaction was performed by the Cross Transformer Fusion module at the high-level semantic stage, where bidirectional attention aligned infrared thermal responses with visible structural cues. Global Attention Mechanism then emphasized target channel and spatial responses. Dynamic adaptive deformable head improved the localization of small, weak, and irregular fire spots. Experiments were conducted on MFire-Bench, a paired visible-infrared benchmark from the public FLAME2-DT, FLAME3, and FireMan-UAV-RGBT datasets. Among them, the dataset also included smoke, low contrast, complex vegetation, and pseudo-high temperature. The available official partitions were retained where provided. In FLAME3, sequence blocks and buffer intervals were used to reduce leakage from adjacent frames. Since the archived data lacked complete, verifiable fire-event or flight identifiers, this partition was interpreted to minimize leakage between adjacent frames, rather than a fully event-independent evaluation. The evaluation provided the precision, recall, mAP@0.5, and mAP@0.5-0.95. A comparison was made between DFF-YOLO and the single-modal and dual-branch baselines under the benchmark protocol. An ablation test was conducted to add CPCA, CTF, GAM, and DAD-Head to the same dual-branch baseline. Paired sensors were then distinguished by the contribution rates of the feature purification, cross-modal alignment, target enhancement, and deformable localization components. The models were trained with the data partition without direct numerical superiority, compared with the different datasets or evaluation protocols. Therefore, the gains were read within the specified benchmark, rather than as a universal estimate of field-deployment performance. DFF-YOLO achieved a precision of 81.3%, a recall of 76.6%, an mAP@0.5 of 81.9%, and an mAP@0.5-0.95 of 51.8%. Compared with the infrared single-modal YOLOv11n and dual-branch DBB-YOLOv11n baselines, the mAP@0.5 increased by 8.7 and 6.6 percentage points, respectively. Ablation tests showed that the dual-branch structure provided the multimodal gain, while CPCA, CTF, GAM, and DAD-Head yielded additional gains in the progressive combination. DFF-YOLO was intended as a candidate for information for fire-scene awareness at the image level, rather than fire warning. Given the task-specific public data, the embedded-platform latency and energy measurements, and GPS/IMU-assisted geolocation, the efficiency can be evaluated on unified hardware using external wildfire data and environmental information. Image detection can be expected for geographic localization and decision-making in future work.