semantic segmentation

In the field of computer vision, semantic segmentation became an important problem that has applications in fields such as autonomous driving and robotics. Image segmentation datasets, on the other hand, present substantial hurdles due to the high intra-class variability, which includes differences across car models or building designs, and the low inter-class variability, which makes it difficult to discern between objects such as buildings that have facades that are visually identical. A focus-enhanced ASPP module that is coupled with an upgraded backbone for semantic segmentation networks is presented in this study as a solution to the problems that have been identified. In order to augment the adaptability of extracted features, the proposed framework utilizes the capability of an attention ASPP module to implement attention processes within the multiscale module. In order to efficiently capture complex features, the encoder stage also makes use of a ResNet-50 backbone that has been properly optimized. In addition, to increase the robustness of the model, data augmentation approaches are applied. mDice of 87.82, mIoU of 79.05, and mean accuracy of 85.2 on the Stanford dataset, and mDice of 88.91, mIoU of 80.03, and mean accuracy of 89.84 on the Cityscapes dataset, according to experimental assessments, demonstrate that the developed technique performs at an accuracy level that is believed to be modern. As a result of these findings, the possibility for greatly improving semantic segmentation performance may be highlighted by integrating attention mechanisms, ASPP modules, and upgraded ResNet structures.

Fire is a major hazard in sensitive environments and can cause irreparable financial and life losses. In addition, fire in the forest and residential areas is considered a threatening event for natural and human resources. Accordingly, detecting fires and smoke in a timely and accurate manner is crucial in preventing financial losses, injuries, and fatalities. Since smoke can be detected before visible flames, smoke detection is a critical component of many fire alarm systems. Sensors sensitive to smoke and fire have the ability to detect these two events, but implementing a huge network of sensors in an open space like a forest is not economical. There are various methods for detecting fire and smoke, and among these, the methods based on deep learning exhibit bigger advantages in terms of accuracy and speed in segmentation. In this paper, we proposed some deep neural networks for fire and smoke detection. These are based on UNet, UNet++, and UNet3+. A proposed FireNet and five other structures are tried as the encoder’s backbone to segment fire and smoke. To train the models, 1200 images gathered from Internet images and videos were prepared, with appropriate labels for smoke and fire applied to their pixels. Experiments show that the best IoU (88.33%) is achieved by UNet++ with EfficientNet.B0 backbone. In small-scale fires, UNet with FireNet has the best performance, and when computational cost is important, UNet3+ with FireNet as the encoder’s backbone is the optimal choice.

Accurate segmentation of lesions from dermoscopic images is very important for timely diagnosis and treatment of skin cancers. Due to the variety of shapes, sizes, colors, and locations of lesions in dermoscopic images, automatic segmentation of skin lesions remains a challenge. In this study, a two-stage method for the segmentation of skin lesions based on deep learning is presented. In the first stage, convolutional neural networks (CNNs) estimate the approximate size and location of the lesion. A sub-image around the estimated bounding box is cropped from the original image. The sub-image is resized to an image of a predefined size. In order to segment the exact area of the lesion from the normal image, other CNNs are used in the DeepLab structure. The accuracy of the normalization stage has a significant impact on the final performance. In order to increase the normalization accuracy, a combination of four networks in the structure of Yolov3 is used. Two approaches are proposed to combine Yolov3 structures. The segmentation results of two networks in the DeepLab v3+ structure are also combined to improve the performance of the second stage. Another challenge is the small number of training images. To overcome this problem, the data augmentation is used, as well as using different modes of an image in each stage. In order to evaluate the proposed method, experiments are performed on the well-known ISBI 2017 dataset. Experimental results show that the proposed lesion segmentation method outperforms the state-of-the-art methods.