This paper presents a novel two-stage image segmentation method using an edge scaled energy functional based on local and global information for intensity inhomogeneous image segmentation. In the first stage, we integrate global intensity term with a geodesic edge term, which produces a preliminary rough segmentation result. Thereafter, by taking final contour of the first stage as initial contour, we begin second stage segmentation process by integrating local intensity term with geodesic edge term to get final segmentation result. Due to the suitable initialization from the first stage, the second stage precisely achieves desirable segmentation result for inhomogeneous image segmentation. Two stage segmentation technique not only increases the accuracy but also eliminates the problem of initial contour existed in traditional local segmentation methods. The energy function of the proposed method uses both global and local terms incorporated with compacted geodesic edge term in an additive fashion which uses image gradient information to delineate obscured boundaries of objects inside an image. A Gaussian kernel is adapted for the regularization of the level set function and to avoid an expensive re-initialization. The experiments were carried out on synthetic and real images. Quantitative validations were performed on Multimodal Brain Tumor Image Segmentation Benchmark (BRATS) 2015 and PH² skin lesion database. The visual and quantitative comparisons will demonstrate the efficiency of the proposed method.

Segmentation of left and right ventricles plays a crucial role in quantitatively analyzing the global and regional information in the cardiac magnetic resonance imaging (MRI). In MRI, the intensity inhomogeneity and weak or blurred object boundaries are the problems, which makes it difficult for the intensity-based segmentation methods to properly delineate the regions of interests (ROI). In this paper, a hybrid signed pressure force function (SPF) is proposed, which yields both local and global image fitted differences in an additive fashion. A characteristic term is also introduced in the SPF function to restrict the contour within the ROI. The overlapping dice index and Hausdorff-Distance metrics have been used over cardiac datasets for quantitative validation. Using 2009 LV MICCAI validation dataset, the proposed method yields DSC values of 0.95 and 0.97 for endocardial and epicardial contours, respectively. Using 2012 RV MICCAI dataset, for the endocardial region, the proposed method yields DSC values of 0.97 and 0.90 and HD values of 8.51 and 7.67 for ED and ES, respectively. For the epicardial region, it yields DSC values of 0.92 and 0.91 and HD values of 6.47 and 9.34 for ED and ES, respectively. Results show its robustness in the segmentation application of the cardiac MRI.

We propose a novel method that combines the strengths of two popular class activation mapping techniques, GradCAM++ and ScoreCAM, to improve the interpretability and localization of convolutional neural networks (CNNs). Our proposed method, called “Grad++-ScoreCAM”, first utilizes the GradCAM++ algorithm to generate a coarse heatmap of an input image, highlighting the regions of importance for a particular class. Then, we employ the ScoreCAM algorithm to refine the heatmap by incorporating the localization information from the intermediate layers of the network. By combining these two techniques, we can generate more accurate and fine-grained heatmaps that highlight the regions of the input image that are most relevant to the prediction of the CNN. We evaluate our proposed method on a benchmark dataset and demonstrate its superiority over existing methods in terms of accuracy and interpretability. Our method has potential applications in various fields, including medical imaging, object recognition, and natural language processing.

Surveillance videos are crucial for crime prevention and public safety, yet the challenge of defining abnormal events hinders their effectiveness, limiting the applicability of supervised methods. This paper introduces an unsupervised end-to-end architecture for video anomaly detection that applies spatial and temporal features to identify anomalies in surveillance footage. The model employs a three-dimensional (3D) convolutional autoencoder, with an encoder-decoder structure that learns spatiotemporal representations and reconstructs the input through the latent space. Skip connections linking the encoder and decoder blocks facilitate the transfer of information across various scales of feature representations, enhancing the reconstruction process and improving the overall performance. The architecture incorporates spatial attention modules that highlight informative regions in the input, enabling improved anomaly detection. Spatial and contextual dependencies are further acquired using 3D convolutional filters. The performance of the proposed model is assessed on four benchmark datasets: UCSD Pedestrian 1, UCSD Pedestrian 2, CUHK Avenue, and ShanghaiTech. Notably, the proposed model achieves frame-based Area Under the Curve (AUC) scores of 94.6% on UCSD Ped 1, 96.7% on UCSD Ped 2, 84.7% on CUHK Avenue, and 74.8% on ShanghaiTech. These results demonstrate the state-of-the-art performance of the proposed approach, highlighting its efficacy in real-world anomaly detection scenarios.

Fashion image analysis has attracted significant research attention owing to the availability of large-scale fashion datasets with rich annotations. However, existing deep learning models for fashion datasets often have high computational requirements. In this study, we propose a new model suitable for low-power devices. The proposed network is a one-stage detector that rapidly detects multiple cloths and landmarks in fashion images. The network is designed as a modification of the EfficientDet originally proposed by Google Brain. The proposed network simultaneously trains the core input features with different resolutions and applies compound scaling to the backbone feature network. The bounding box/class/landmark prediction networks maintain the balance between the speed and accuracy. Moreover, a low number of parameters and low computational cost make it efficient. Without image preprocessing, we achieved 0.686 mean average precision (mAP) in the bounding box detection and 0.450 mAP in the landmark estimation on the DeepFashion2 validation dataset with an inference time of 42 ms. We obtained optimal results in extensive experiments with loss functions and optimizers. Furthermore, the proposed method has the advantage of operating in low-power devices.