Heterogeneous Gray-Temperature Fusion-Based Deep Learning Architecture for Far Infrared Small Target Detection

Figure 6

Performance comparison between the normalization and pixel-level fusion method. (a)-(e) are the results for the winter test set, and (f)-(j) are the results for the summer test set. The first row was normalized with the mean and standard deviation of 0.5; the second row is between 0 and 1; the third row is between -1 and 0; and the fourth row is between -1 and 1. The last row uses the precomputed mean and standard deviation for large scale dataset. (a)-(j) are the performance results according to the normalization method and fusion method, and (k) and (l) were obtained by collecting only the best performance of each normalization method for the summer and winter test set, respectively.