An Automated Diabetic Retinopathy Lesion Segmentation Using Dilated Convolution patch Merging Swin transformer G. Indira Devi, D. Madhavi 2025 5th Asian Conference on Innovation in Technology Asiancon 2025, 2025 Diabetic retinopathy (DR) lesions segmentation is crucial for identification of DR. The main problems facing with this task is that lesions of DR may have different shapes, sizes, brightness and color. Manual approaches may not give efficient results. In order to overcome the limitations with the manual approaches, this paper propose a segmentation based approach which detects and segments four typical DR lesions Microaneurysms(MA), Hard Exudates(HE), Hemorrhages(HM) and Neovascularization simultaneously based on the method of Deep Learning. The raw image is first pre-processed and then segmented using Dilated Convolutional Patch Merging Swin Transformer. Experiment was conducted on two public datasets including APTOS-2019 and IDRiD. The results of the IoU, Dice, sensitivity and specificity show that our approach achieves competitive performance. The proposed approach achieved 99.4% accuracy, 98.6% precision, 98.1% sensitivity, 99% F1-score, 99.5% specificity, 0.00375 FPR, 0.015 FNR, 0.014 FDR, and 98.1% MCC on APTOS-2019 dataset. The performance improvement indicates that this approach gives good results for DR lesion detection and segmentation and it also beneficial in other types of segmentation tasks.
An Automated Diabetic Retinopathy Severity Level Classification using Vertically Stacked Squeeze Excitation Network with the Crayfish Optimization Algorithm G.Indira Devi, D. Madhavi 2025 6th International Conference for Emerging Technology Incet 2025, 2025 Diabetic retinopathy (DR) is a serious eye disease that can cause vision loss and blindness. This study proposes an advanced deep learning based DR classification approach for determining DR severity. Input images are collected from the APTOS-2019 dataset. Here, first images are pre-processed using adaptive Gaussian filter and then to increase quality of the images using the Modified Contrast Limited Interval-value Fuzzy Histogram Equalization (MCLI-FHE) method. Then, the image is segmented using a Patch Merging Swin Transformer (PMST). After segmentation, the important features get extracted using the Gray Level Dependence Matrix (GLDM) and Gray Level Co-occurrence Matrix (GLCM). Subsequently, the DR grouping is performed using Vertically Stacked Residual Convolutional Long short-term Enhanced Crayfish Squeeze Excitation Network (VSRCL-ECSEN). The Crayfish Optimization Algorithm (COA) is used for hyper parameter tuning. As a result, the proposed model accurately detects the corresponding lesions and classifies DR as mild, moderate, severe, proliferative, or no DR. For the APTOS-2019 dataset, the proposed approach achieved 99.4% accuracy, 98.6% precision, 98.1% sensitivity, 99% F1-score, 99.5% specificity, 0.00375 FPR, 0.015 FNR, 0.014 FDR, and 98.1% MCC results.
An Optimized Deep Features for Detecting Tampered Region from the Copy Move Forgery Image Allu Venkateswara Rao, D. Madhavi International Journal of Engineering Trends and Technology, 2024 Nowadays, forgery detection systems have rapidly grown in the digital application to find crime events. However, detecting the forgery and identifying the forged tampered portion is more complex because of the noisy data. To overcome this issue, the current research article has aimed to develop a novel Lion-based Optimized Radial Basis Neural Model (LORBNM). Initially, the CoMoFoD dataset has been trained, the training noise has been removed from the pre-processing layer, and then the error-free images are entered into the classification layer. Consequently, the classification parameters were tuned, and the present features were extracted. Furthermore, the image types have been specified in terms of Computer-Generated-Image (CGI), Natural Image (NI), and Forgery Image (FI). Eventually, the tapered region was predicted and segmented from the forgery image, and then the key metrics were calculated and compared with other existing approaches. In that, the presented LORBNM has observed the finest segmentation exactness score.
RCVM-ASS-CICSKA-PAPT-VDF: VLSI design of high-speed reconfigurable compressed Vedic PAPT-VDF filter for ECG medical application K. V. Suresh Kumar, D. Madhavi Transactions on Emerging Telecommunications Technologies, 2024 During signal acquisition, the signals are impacted by multiple noise sources that must be filtered before any analysis. However, many different filter implementations in VLSI are dispersed among many studies. This study aims to give readers a systematic approach to designing a Pipelined All‐Pass Transformation based Variable digital filter (PAPT‐VDF) to eliminate the high‐frequency noise from ECG data. The modified design emphasizes first‐ and second‐order responses to obtain high‐speed filter realization with high operating frequencies. The addition of adder and multiplier designs to the hardware architecture of a filter design improves performance. The fundamental blocks of the filter design are the adder and multiplier. The adder and multiplier are employed with an Adaptable stage size‐based concatenation, incremented carry‐skip adder (ASS‐CICSKA), and Improved reconfigurable compressed Vedic multiplier (IRCVM). Utilizing the adder design diminishes the delay with enhanced performance because receiving the carry from an incrementation block is not mandatory. In the multiplier design, the compressor and the reconfigurable approach are adapted with a data detector block to detect the redundant input and lower the logic gates' switching activity with less area overhead. The proposed filter design is implemented in vertex 7 FPGA family device, and the performance measures are analyzed regarding area utilization, delay, power, and frequency. Also, by using the denoised signal, the mean square error (MSE), and signal‐to‐noise ratio (SNR) are evaluated in the MATLAB platform.
Unsupervised Content Mining in CBIR: Harnessing Latent Diffusion for Complex Text-Based Query Interpretation Venkata Rama Muni Kumar Gopu, Madhavi Dunna Journal of Imaging, 2024 The paper demonstrates a novel methodology for Content-Based Image Retrieval (CBIR), which shifts the focus from conventional domain-specific image queries to more complex text-based query processing. Latent diffusion models are employed to interpret complex textual prompts and address the requirements of effectively interpreting the complex textual query. Latent Diffusion models successfully transform complex textual queries into visually engaging representations, establishing a seamless connection between textual descriptions and visual content. Custom triplet network design is at the heart of our retrieval method. When trained well, a triplet network will represent the generated query image and the different images in the database. The cosine similarity metric is used to assess the similarity between the feature representations in order to find and retrieve the relevant images. Our experiments results show that latent diffusion models can successfully bridge the gap between complex textual prompts for image retrieval without relying on labels or metadata that are attached to database images. This advancement sets the stage for future explorations in image retrieval, leveraging the generative AI capabilities to cater to the ever-evolving demands of big data and complex query interpretations.
Zero-Shot Sketch-Based Image Retrieval Using StyleGen and Stacked Siamese Neural Networks Venkata Rama Muni Kumar Gopu, Madhavi Dunna Journal of Imaging, 2024 Sketch-based image retrieval (SBIR) refers to a sub-class of content-based image retrieval problems where the input queries are ambiguous sketches and the retrieval repository is a database of natural images. In the zero-shot setup of SBIR, the query sketches are drawn from classes that do not match any of those that were used in model building. The SBIR task is extremely challenging as it is a cross-domain retrieval problem, unlike content-based image retrieval problems because sketches and images have a huge domain gap. In this work, we propose an elegant retrieval methodology, StyleGen, for generating fake candidate images that match the domain of the repository images, thus reducing the domain gap for retrieval tasks. The retrieval methodology makes use of a two-stage neural network architecture known as the stacked Siamese network, which is known to provide outstanding retrieval performance without losing the generalizability of the approach. Experimental studies on the image sketch datasets TU-Berlin Extended and Sketchy Extended, evaluated using the mean average precision (mAP) metric, demonstrate a marked performance improvement compared to the current state-of-the-art approaches in the domain.
Identification of Microaneurysms and Exudates for Early Detection of Diabetic Retinopathy G Indira Devi, D. Madhavi International Journal of Advanced Computer Science and Applications, 2023 —Diabetic retinopathy (DR) is a condition that may be a complication of diabetes, and it can damage both the retina and other small blood vessels throughout the body. Microaneurysms (MA’s) and Hard exudates (HE’s) are two symptoms that occur in the early stage of DR. Accurate and reliable detection of MA’s and HE’s in color fundus images has great importance for DR screening. Here, a machine learning algorithm has been presented in this paper that detects MA’s and HE’s in fundus images of the retina. In this research a dynamic thresholding and fuzzy c mean clustering with characteristic feature extraction and different classification techniques are used for detection of MA’s and HE’s. The performance of system is evaluated by computing the parameters like sensitivity, specificity, accuracy, and precision. The results are compared between different types of classifiers. The Logistic Regression classifier (LRC) performance is good when compared with other classifiers with an accuracy of 94.6% in detection of MA’s and 96.2% in detection of HE’s.
Stacked Siamese Neural Network (SSiNN) on Neural Codes for Content-Based Image Retrieval Gopu V. R. Muni Kumar, D. Madhavi IEEE Access, 2023 Content-based image retrieval (CBIR) represents a class of problems that aims at finding relevant images in response to an image-based search query. The CBIR systems use similarity measures or distance metrics between a group of representative features in the query image and those in the image repository. Traditionally, these features were generated by hand, employing image features such as colour, texture, shape, and so on. Due to the fact that these methods do not provide a comprehensive perspective of the images, they cannot be widely utilized in contemporary CBIR systems. This is due to the so-called semantic gap between query intent and system perspective. The most recent advancements in deep learning offer a viable alternative to manually built features, leveraging the representational learning capability of deep neural networks. This paper presents a method of implementing a CBIR system using a multi-stage approach known as classify, differentiate, and retrieve (CDR). The first stage involves using a deep neural network to encode the images. Later, a custom-trained stacked Siamese Neural network (SSiNN) is employed to differentiate the latent space representation of the images obtained from the first stage. The experimental results for the CIFAR-10 dataset were presented, along with an algorithm for applying this strategy to any generic dataset. Experimental outcomes demonstrate that the proposed strategy is superior to the current best practices.
OPTIMIZATION of QBIC SYSTEMS USING LG WAVELETS International Journal of Interdisciplinary Global Studies, 2020
High performance architectures for recursive loop algorithms 2009 International Conference on Control Automation Communication and Energy Conservation Incacec 2009, 2009
