mohammad javad fadaeieslam

Thermal imaging systems, due to their unique features, such as the ability to record images in different weather conditions, recording images at night, or having anti-counterfeiting properties, have special military, security and judicial applications. However, the images recorded by thermal cameras cannot be recognized by the human eye, and it is very difficult for humans to recognize the faces of thermal images. Converting thermal images to visible light images is in the field of image-to-image content transfer or image to image translation. So far, many deep learning models have been introduced to convert thermal images into visible light. Among these models, adversarial networks have been able to achieve significant progress in this field. In this paper, an attempt was made to improve the ClawGAN network, which is specifically designed to convert thermal images into visible light. Our method is based on the integration of effective techniques such as Unet++, Unet3+, self-attention network in the  generator of the base model. In this way, the network will be able to produce higher quality images that can be recognized by the human eye and have minimal distortion, blur and noise when transferring content from the thermal domain to the visible light domain. The obtained results showed that the proposed generator was able to significantly improve the evaluation criteria such as MSE, PSNR, RMSE, UQI and PSNR-B.

The image-to-image translation is one of the most challenging topics in artificial intelligence, which has recently made significant progress with the use of generative adversarial networks (GANs). However, existing methods often fail to translate the noise source to the target domain. This article presents the WTGAN network, which includes a new generator and a local and global discriminator to solve this problem. The generating network is designed based on wavelet transform and attention. Due to the fact that wavelet transforms are powerful tools for removing general noise from the image, They have been used in the structure of the generator. Also, attention, residual and skip-connections can provide deeper surface information between the source and target image and help to improve the generator performance. Experiments were performed on the Cityscapes dataset and PSNR, SSIM, and LPIPS criteria were used for evaluation. The results have shown that the model can well reduce the effects of noise at the source, well reserve structure, and achieve the desired quality.

Detection of salient objects is done with the aim of identifying and segmenting prominent objects or areas in an image. Fully Convolutional Networks (FCNs) have shown their advantages in salient object detection; however, many previous works have focused on the accuracy of the prominent area without paying attention to its edge. This paper focuses on the complementarity between edge information and salient object one and added an edge recognition module to explicitly model edge information to maintain salient object boundaries. Our proposed network is trying to improve these two tasks simultaneously. The presence of objects with different scales in related datasets is another problem in this area. It requires an appropriate cost function to deal with the imbalance problem between background and foreground in images. So, we have used the hybrid cost function in the training phase, which is not sensitive to the scale of objects and can better manage the problem of spatial coherence and uniformly highlight salient areas without additional parameters. A Comparison of the quantitative and qualitative results obtained by the proposed method with other advanced methods in six widely used protrusion detection datasets shows that the proposed method has a good performance and can quickly identify prominent areas. In particular, according to the quantitative results, our method gets the best result on three widely used test datasets in terms of F-measure and MAE criteria, demonstrating the proposed method's efficiency.

Detecting objects most focused on by human eyes when viewing a scene is of great interest to the computer vision community. Although a large amount of detection algorithms are available, due to variety and complexity of the image structures, the obtained saliency maps are still not satisfying enough. In this paper, we have proposed an efficient, supervised algorithm for saliency map detection which uses a conditional random field. Integrating different salient cues with matrix decomposition methods through CRF is one of the innovations of this paper. Another achievement of this paper is considering potential weights, obtained from CRF training process, as a ranking tool to select the best saliency cues. Since CRF is a supervised method, some papers select, for training step, a number of images which are most similar to the input image. The present paper offers, as our third contribution, a comprehensive assessment of the methods which select such similar images. Evaluating the proposed method on the ECSSD and MSRA-10k datasets with respect to the evaluation criteria has indicated its excellent performance.

In this paper, a two stage co-segmentation method,based on matrix decomposition,has been proposed. In the first step, each image is segmented into some super-pixels and then salient parts of each image are extracted via structured matrix decomposition (SMD) method. The low-rank matrix represents image background and the sparse matrix contains salient objects. In this step, the super-pixels that are partitioned as background with high confidence will be removed. In the second step, all remaining super-pixels are considered all together and the tree structure is rearranged and then the SMD method is applied again to this new data. Parts of the common salient object compose the low-rank matrix due to the large number of them in the remaining super-pixels. In other words, the proposed approach has embedded intra-image adjacency information and inter-images similarity information into the matrix decomposition method via proper weighting of the tree structure.iCoseg dataset has been used to evaluate its performance. The results demonstrate its effectiveness and superiority.

One of the most important processing steps in the human vision system is the detection of a scene saliency map. Since saliency map can be applied to algorithms such as segmentation, compression and image retrieval, Researchers have focused on providing an efficient model to recognize it. Although a lot of works have been done in this area, the obtained saliency maps are still not satisfying enough. For this purpose, we propose a simple and supervised algorithm to identify the saliency map using a conditional random field (CRF) and saliency cues. In the proposed method, local contrast, center-bias, and backgroundness features have been used for CRF training. Additionally, a new feature based on matrix decomposition has been employed to improve the performance. In the following, CRF has been trained according to the features of 20 images close to the input image. Finally, input image saliency is estimated according to calculated weights in the training phase, input image saliency cues, and ground truths. The proposed method outperforms other methods in terms of algorithm implementation accuracy and speed.