Representation Mimicking Loss for Dehazing: A Feature Matching and Haze Density Aware Approach

Representation Mimicking Loss. As the intermediate representations of the clean image extracted by Teacher have abundant knowledge for reconstruction. Mimicking the feature distribution of the clean images might be helpful for training the dehazing network. Hence, we introduce two different types of representation mimicking loss: feature matching imitation loss (FMIL) and haze density aware imitation losses which weights FMIL (WFMIL). For simplicity, we let T'(J) denotes the feature maps of the mth layer of Teacher trained on clean images, let S'(I) denotes the feature maps of the nth layer of Student trained on hazy images, and let g be a linear transformation function such as a pointwise convolution which aims to make sure the number of channels of T'(J) and S'(I) is consistent.

FMIL is formulated as: xxxxx, where C is a set of candidate pairs of feature representations.

Since dense hazy regions are seriously degraded caused by the uneven concentration of haze in an image, the dehazing network often fails in dehazing well in the earlier stage. To overcome this issue, we design a haze density aware imitation loss to up-weight the dense hazy region. Specifically, we propose to use the concentration of haze to weight the representation mimicking loss. In our solution, we employ a transmission map and the residual between the hazy image and ground-truth in gray to measure the concentration of haze. As shown in Figure 5 (c), the transmission map is highly related to the scene depth and provides smooth guidance because it ignores the details of objects. However, the transmission maps are not always available in the training set. Hence, we measure the concentration of haze by calculating the residual between the hazy image and ground-truth in gray, as shown in Figure 5 (d). With these two measurements, WFMIL is formulated as: vvvv.

where t is the transmission map obtained from the training set, Ig and Jg are the gray image of I and J, respectively. norm() denotes a normalization operator which makes the pixel values normalized to [0,1]. Specifically, for a given xxxx, where min(I) denotes the minimum value of I, max(I) denotes the maximum value of I, (i, j) is the pixel location.