In recent years, Convolutional Neural Networks (CNNs) have made significant strides in the field of Segmentation, particularly in Semantic Segmentation where both accuracy and efficiency are crucial. However, despite their high accuracy, these deep networks are not practical for real-time use due to their low inference speed. This issue has prompted researchers to explore various techniques to improve the efficiency of CNNs. One such technique is knowledge distillation, which involves transferring knowledge from a larger, cumbersome (teacher) model to a smaller, more compact (student) model. This paper proposes a simple yet efficient approach to address the issue of low inference speed in CNNs using knowledge distillation. The proposed method involves distilling knowledge from the feature maps of the teacher model to guide the learning of the student model. The approach uses a straightforward technique known as pixel-wise distillation to transfer the feature maps of the last convolution layer of the teacher model to the student model. Additionally, a pair-wise distillation technique is used to transfer pair-wise similarities of the intermediate layers. To validate the effectiveness of the proposed method, extensive experiments were conducted on the PascalVoc 2012 dataset using a state-of-the-art DeepLabV3+ Segmentation network with different backbone architectures. The results showed that the proposed method achieved a balanced mean Intersection over Union (mIoU) and training time.

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.

Recent advancements in Weakly Supervised Semantic Segmentation (WSSS) have highlighted the use of image-level class labels as a form of supervision. Many methods use pseudo-labels from class activation maps (CAMs) to address the limited spatial information in class labels. However, CAMs generated from Convolutional Neural Networks (CNNs) are often led to focus on prominent features, making it difficult to distinguish foreground objects from their backgrounds. While recent studies show that features from Vision Transformers (ViTs) are more effective in capturing the scene layout than CNNs, the use of hierarchical ViTs has not been widely studied in WSSS. This work introduces "SWTformer" and explores the effect of Swin Transformer’s local-to-global view on improving the accuracy of initial seed CAMs. SWTformer-V1 produces CAMs solely based on patch tokens as its input features. SWTformer-V2 enhances this process by integrating a multi-scale feature fusion mechanism and employing a background-aware mechanism that refines the accuracy of localization maps, resulting in better differentiation between objects. Experiments on the Pascal VOC 2012 dataset demonstrate that compared to state-of-the-art models, SWTformer-V1 achieves 0.98% mAP higher in localization accuracy and generates initial localization maps that are 0.82% mIoU higher in accuracy while relying solely on the classification network. SWTformer-V2 enhances the accuracy of the seed CAMs by 5.32% mIoU. Code available at: ttps://github.com/RozhanAhmadi/SWTformer

This paper presents a Semantic method for aerial image Segmentation. Multi-class aerial images are often featured with large intra-class variations and inter-class similarities. Furthermore, shadows, reections and changes in viewpoint, high and varying altitude and variability of natural scene pose serious problems for simultaneous Segmentation. The main purpose of Segmentation of aerial images is to make subsequent recognition phase straightforward. Present algorithm combines two challenging tasks of Segmentation and classification in a manner that no extra recognition phase is needed. This algorithm is supposed to be part of a system which will be developed to automatically locate the appropriate site for Unmanned Aerial Vehicle (UAV) landing. With this perspective, we focused on segregating natural and man-made areas in aerial images. We compared different classifiers and explored the best set of features for this task in an experimental manner. In addition, a certainty based method has been used for integrating color and texture descriptors in a more efficient way. The experimental results over a dataset comprised of 25 high-resolution images show the overall binary Segmentation accuracy rate of 91.34%.