About
Chanchal Ghosh is a researcher in Computer Science at Independent Researcher. Based in India. Research interests include deep learning, image processing.
Research keywords
Publications
9A hybrid deep fuzzy clustering framework with fusion-based learning for robust SAR change detection
Multi-temporal Synthetic Aperture Radar (SAR) images are essential in detecting changes in the environment, analyzing urban growth, and assessing disasters. Nonetheless, it is difficult to reliably detect meaningful changes because of speckle noise, radiometric variations, and inhomogeneous terrain features that severely impair the accuracy of the detection. This paper presents a hybrid change detection framework that integrates Fusion-based Difference Imaging (FDI) with a deep feature-guided clustering method. The proposed FDI module enhances sensitivity to subtle backscatter variations, which combines complementary difference representations to enhance the discriminability of changed and unchanged regions. These representations are refined through a deep clustering process that improves separability and reduces false alarms. Extensive experiments on benchmark multi-temporal SAR data sets indicate that the proposed approach has high detection accuracy and robustness with overall accuracy (PCC) of up to 99.97%, a Kappa coefficient of up to 98.30% and an F1-score of up to 99.59% across various datasets. The proposed approach consistently outperforms competing methods across evaluated datasets and different speckle conditions. The results show that fusion-based representation learning, with uncertainty-aware clustering, is a scalable, efficient, and flexible solution to robust SAR change detection in challenging imaging conditions.
Hierarchical Deep Learning Ensemble Framework for Multi-Class Rice Foliar Disease Diagnosis: A Comparative Architecture Analysis
Infestations of foliar diseases in rice plants are common and can reduce harvest yields and affect food supplies worldwide. A system for the automatic detection of these diseases was developed in this study using seven different deep learning models. Six common types of rice leaf diseases were tested using models such as EfficientNet (B0 and B7), ResNet50, InceptionV3, VGG16, and VGG19. The proposed framework integrates the advantages of all models, assigning greater significance to those that exhibit superior performance. By achieving 96.97% accuracy while retaining speed and lightweight features, MobileNetV2 demonstrated superior performance. Both InceptionV3 and EfficientNetB7 performed well, reporting accuracies of 96.78% and 96.40%, respectively. It was also observed that newer, more efficient models exhibited markedly superior performance compared to older deep networks. This method makes it easier to bridge the gap between the urgent need for rapid disease detection on farms and the lack of agricultural experience. The system, which uses low-cost equipment, helps small farmers all over the world diagnose diseases accurately, resulting in better yields of crops.
Dynamic Attention-guided Heterogeneous Mayfly Optimization for Hierarchical Swin Transformer-based Rice Leaf Disease Recognition-demo
Overview This repository contains a Python-based deep learning implementation of a Swin Transformer model integrated with a Heterogeneous Mayfly Optimization (HMO)-inspired hyperparameter tuning strategy for automated rice leaf disease classification. The framework aims to improve classification performance by optimizing important training hyperparameters, including learning rate, weight decay, and dropout rate, using validation accuracy as the fitness function. Methodology The proposed implementation consists of the following stages: Dataset Preparation: Rice leaf disease images are loaded using PyTorch ImageFolder, resized to 224 × 224 pixels, normalized, and divided into training (80%) and validation (20%) subsets. Swin Transformer Architecture: A pretrained Swin Transformer Base model (swin_base_patch4_window7_224) is employed for multi-class rice leaf disease classification. HMO-Inspired Hyperparameter Optimization: A simplified Heterogeneous Mayfly Optimization strategy searches for suitable learning rate, weight decay, and dropout configurations. Each candidate is evaluated using validation accuracy after a short training phase. Final Model Training: The Swin Transformer is trained using the optimized hyperparameters and the AdamW optimizer with cross-entropy loss. Performance Evaluation: Classification performance is evaluated using validation accuracy, classification reports, and confusion matrices. Implementation Details Deep learning framework: PyTorch Model architecture: Swin Transformer Base Pretrained model weights: ImageNet-pretrained weights Input image resolution: 224 × 224 Batch size: 16 Optimizer: AdamW Loss function: Cross-Entropy Loss Hyperparameter optimizer: HMO-inspired optimization Population size: 4 Optimization iterations: 3 Hyperparameters optimized: Learning rate, weight decay, and dropout rate Final training epochs in the supplied notebook: 5 Intended Application The implementation is intended for research and experimentation in automated plant disease recognition, agricultural image classification, and metaheuristic-based deep learning hyperparameter optimization. Reproducibility and Availability The repository provides the Jupyter Notebook containing the model implementation, dataset loading procedure, hyperparameter optimization algorithm, training pipeline, and evaluation functions. The rice leaf disease dataset is not included in this upload and must be obtained separately. The dataset directory path should be configured according to the user's local or cloud environment before execution. This software is provided for academic and research purposes.
Detection of Leukemia Using Transfer Learning
Leukemia is a kind of cancer that targets White Blood Cells (WBCs) and influences equally children and adults, caused by the neoplastic proliferation of bone marrow precursor cells, resulting in early death and a variety of other side effects. As a result, various manual techniques for detecting leukemia have been developed, but their performance is not up to the mark as the results were not accurate and reliable. A bone marrow investigation may be advised to check and determine the kind of blood cancer or leukemia. This conventional method is time-taking, and the accuracy of detection depends upon the efficiency of the pathologists. Early detection by using computerized techniques can significantly enhance the rate of cure. Convolutional Neural Networks (CNNs) are more and more getting used in the classification and diagnosis of medical image processing. This article suggests a new automated classification technique using transfer learning to identify Leukemia based on microscopic blood images, which will overcome the drawbacks of traditional methods and provides more accuracy. Pre-processed blood smear images, move on to feature extraction. In the feature extraction phase, Inception V3, a pre-trained Deep Convolution Neural Network (DCNN) is used to extract features. To shorten training time and improve efficiency, a pre-trained model is applied along with small datasets into a new model. The Inception-V3 model serves as a base, where on top there is a Fully-Connected (FC) layer for optimizing the process of classification. Convolution layers can learn enough with their convolution kernel to produce tensor outputs during the model-building process. Studies were carried out on a dataset of 15135 images, affirming that the model outperforms existing methods due to its 97.34% accuracy in classification.
Deep Learning-Based Transfer Learning for the Detection of Leukemia
Leukemia is typically diagnosed by using a microscope to examine blood and bone marrow smears as well as extremely challenging procedures like the Cytochemical test, which can be used to identify and categorize leukemia. However, these techniques are incredibly pricy, dependent, and time-consuming on the skill and cunning of the experts involved. In contrast to conventional methods, which have several drawbacks, this research suggests a classification model depends upon blood microscopic images for detecting leukemia by utilizing transfer learning. The suggested study demonstrates a computer-aided diagnosis method that uses Mobilenet V2 and Convolutional Neural Networks (CNNs) to identify leukemia images over healthy images. When it comes to using deep learning for image analysis and the comparison outcomes of the current study for diagnostic tasks, the implementation of Mobilenet V2 is relatively simple. Results demonstrate that the suggested model, having a general accuracy of 96.58%, a sensitivity of 95.17%, and a specificity of 98.58%, gives better predictions than separate models for classifying leukemia.
Detection of changes in synthetic aperture radar images using Modified Gauss-Log ratio and Fuzzy Local Information C-Means clustering
A deep learning‐based SAR image change detection using spatial intuitionistic fuzzy C‐means clustering
With the rapid progress of technologies in the arena of remote sensing and satellite imagery, Synthetic Aperture Rader (SAR) images have become an important source of data for research concerning changed detection. Out of the numerous techniques and approaches available for change detection in a particular location, most of them are initially targeted toward producing the difference image. In this article, a change detection approach is suggested that produces the result without finding a difference image. We are motivated toward the design of such an approach to reduce the effect of the difference image. In this method, we generate the training and testing sample for CNN classification directly from the original SAR image without any pre‐processing operations. This reduces the effect of the difference image on the final classification result. Since traditional fuzzy c means (FCM) are highly susceptible to noise and do not give desired results, we use spatial fuzzy c means (sFCM) with an intuitionistic approach. The intuitionistic approach refers to the degree of hesitation resulting from a lack of information. The approach is less hampered by noise and yields better outcomes. The basic idea of this method is to find false levels using spatial intuitionistic fuzzy c means clustering. Thereafter, the CNN is trained using the samples that are selected from the original samples using false labels. Finally, the classification results are generated through the trained CNN. Investigational results using the proposed approach show promising results when applied over two SAR image datasets compared to several existing methodologies.
SAR Image Change Detection Using Modified Gauss-Log Ratio Operator and Convolution Neural Network
Decimal Sequence based Imperceptible Audio Watermarking
Digital watermarking has been proposed as a new, alternative method to enforce intellectual property rights and protect digital media from tempering.Digital watermarking is defined as imperceptible, robust and secure communication of data related to the host signal, which includes embedding into and extracted from the host signal.The main challenge in digital audio watermarking is that if the perceptual transmitter is fixed, the design of watermark system cannot obtain high watermark data rate and high robustness at the same time.The main objective of this paper is to development of an audio watermarking algorithm which obtain high watermark data rate and high quality of audio at the same time General