Gangyi Ding

Fire detection has always been an important issue in the field of emergency. Traditional fire detection methods have many shortcomings, such as high hardware cost and limited application scenarios. This research constructs a detection process based on YOLOv10n models and creates corresponding multi-granularity fire datasets. The accuracy of fire and smoke detection on embedded devices reaches 66.58%. The multi-granularity detection model, based on quantization methods such as KL, ACIQ and EQ, is constructed to distinguish multiple fire and smoke types such as more smoke and less fire, less smoke and more fire, and average smoke and fire. Compared to a single quantization method, the average accuracy increases by 2.75%. The multi-granularity quantization model is verified on embedded devices, and a simulation system based on this model is implemented. The system distributed structure allows monitoring personnel to remotely interact with the model parameters, performing visual fire simulation to better adapt to various application scenarios.

Texture classification is a challenging and pivotal task in the field of computer vision. Recently, methods that extract features through a frozen backbone have shown great potential due to the growing computing cost of large models. However, these methods overlook the interactional information among activation maps at varying depths. To address this issue, we propose Attention Based Random Encoded Activation Maps(AREAM), which utilizes the Convolutional Block Attention Module(CBAM) to extract richer texture information from activation maps at different depths. After processing the activation maps through random auto-encoders, we employ Kernel Principal Component Analysis(KPCA) to eliminate less significant information and reduce dimension. Subsequently, the data is fed into a non-linear Support Vector Machine(SVM) for classification. The experimental results across multiple texture datasets demonstrate the effectiveness of our method.

Deep generative models have made great progress in synthesizing images with arbitrary human poses and transferring the poses of one person to others. Though many different methods have been proposed to generate images with high visual fidelity, the main challenge remains and comes from two fundamental issues: pose ambiguity and appearance inconsistency. To alleviate the current limitations and improve the quality of the synthesized images, we propose a pose transfer network with augmented D isentangled F eature C onsistency (DFC-Net) to facilitate human pose transfer. Given a pair of images containing the source and target person, DFC-Net extracts pose and static information from the source and target respectively, then synthesizes an image of the target person with the desired pose from the source. Moreover, DFC-Net leverages disentangled feature consistency losses in the adversarial training to strengthen the transfer coherence and integrates a keypoint amplifier to enhance the pose feature extraction. With the help of the disentangled feature consistency losses, we further propose a novel data augmentation scheme that introduces unpaired support data with the augmented consistency constraints to improve the generality and robustness of DFC-Net. Extensive experimental results on Mixamo-Pose and EDN-10k have demonstrated DFC-Net achieves state-of-the-art performance on pose transfer.