Research On Fine-grained Classification Of Flame Images Of Sintering Based On Multi-scale Heterogeneous CNN

In the metallurgical industry,stable and high-quality sintering production process is an important guarantee for blast furnace smelting.The characteristic information of the sintering production process and the sintering end state is contained in the sintering flame image.Accurate and effective flame area tracking and fine-grained classification of sintering flame images is helpful for analyzing sintering process state and sintering end state,providing timely and reliable guidance information for the actual sintering production process,and has strong practical significance in engineering.In view of the existence of a large number of smoke and dust particles in the sintering flame image acquisition environment,the image was subjected to anisotropic diffusion denoising and enhancement to weaken the background features and enhance the feature expression ability of the image.Based on the particle filter algorithm,the RGB color space of the sintering flame image was introduced as the feature of particle state transition,and the particle weight and resampling strategy were optimized to solve the problem of particle degradation and sample poverty,so as to realize the effective tracking of the flame area of the sintered cross-section image,and then analyzed sintering process status.Comparing the effect of the classical convolutional neural network model on the dataset,Inception-v3 was selected as the backbone network.Further use of transfer learning methods and regularization processing,accelerated the iterative convergence of the model while preventing the model from overfitting.The experimental results showed that the improved particle filter algorithm tracks the flame area pixel area more than 90%,and the tracking effect is good and stable.The multi-scale convolutional neural network model based on Inception-v3 achieved 98.91% accuracy in fine-grained classification.The flame area tracking and fine-grained classification of the sintering flame image provided a theoretical basis for analyzing the state of the sintering process and the end state,and effectively improved the efficiency and quality of the sintering production.Figure 34;Table 8;Reference 56...