International Journal of Image, Graphics and Signal Processing (IJIGSP)

IJIGSP Vol. 16, No. 4, Aug. 2024

Cover page and Table of Contents: PDF (size: 663KB)

Table Of Contents

REGULAR PAPERS

Image Analysis of Impurity in Machine-harvested Cotton Based Machine Vision

By Mingjie LI Vladimir Y. Mariano

DOI: https://doi.org/10.5815/ijigsp.2024.04.01, Pub. Date: 8 Aug. 2024

The mechanization rate of cotton picking continues to increase with the continuous improvement and development of China's agricultural modernization level. However, when picking cotton, the machine cannot distinguish between cotton fibers and impurities well, resulting in a certain gap in impurity content compared to manually picked cotton. This paper combines machine vision and image processing technology to adopt an improved Canny-based impurity image processing algorithm. By performing light processing, selecting a color space, filtering images, and removing noise from machine-harvested cotton images, the suppression of virtual edges on impurity images allows for more accurate identification of impurities on the cotton surface. Finally, experimental details and results conclusively demonstrate the effectiveness of this method, providing a basis for detecting and classifying cotton impurities.

[...] Read more.
CRDL-PNet: An Efficient DeepLab-based Model for Segmenting Polyp Colonoscopy Images

By Anita Murmu Piyush Kumar Shrikant Malviya

DOI: https://doi.org/10.5815/ijigsp.2024.04.02, Pub. Date: 8 Aug. 2024

Colorectal cancers are the third-largest kind of cancer in the world. However, detecting and removing precursor polyps with adenomatous cells using optical colonoscopy images helps to prevent this type of cancer. Moreover, hyperplastic polyps are benign cancers; adenomatous polyps are more likely to grow into cancerous tumors. Therefore, the detection and segmentation of polyps provide further histological evaluation. However, the main challenge is the extensive range of infected polyp features inside the colon and the lack of contrast between normal and infected areas. To solve these issues, the proposed novel Customized ResNet50 with DeepLabV3Plus Network (CRDL-PNet) model provided a scheme for segmenting polyps from colonoscopy images. The customized ResNet50 extracted features from polyp colonoscopy images. Furthermore, Atrous Spatial Pyramid Pooling (ASPP) is used to handle scale variation during training and improve feature selection maps in an upsampling layer. Additionally, the Gateaux Derivatives (GD) approach is used to segment boundary boxes of polyp regions. The proposed method has been evaluated on four datasets, namely the Kvasir-SEG, ETIS-PolypLaribDB, CVC-ClinicDB, and CVC-ColonDB datasets, for segmenting and detecting polyps. The simulation results have been examined by evaluation metrics, such as accuracy, Intersection-Over-Union (IOU), mean IOU, precision, recall, F1-score, dice, Jaccard, and Mean Process Time per Frame (MPTF) for proper validation. The proposed scheme outperforms the existing State-Of-The-Arts (SOTA) model on the same polyp datasets.

[...] Read more.
Seamless Panoramic Image Stitching Based on Invariant Feature Detector and Image Blending

By Megha V. Rajkumar K. K.

DOI: https://doi.org/10.5815/ijigsp.2024.04.03, Pub. Date: 8 Aug. 2024

Image stitching is the method of creating a composite image from several images of the same scene. This paper addresses the issues of generating a seamless panoramic image from a series of photographs of the same scene by varying scale, orientation and illumination. A feature-based approach is proposed in this paper. Scale Invariant Feature Transform (SIFT) is used to detect key points in the image. SIFT is both a feature detector and descriptor. The common region between different images is identified by comparing the feature descriptors of each image. Brute-Force matcher with KNN algorithm is used for feature matching. The outliers in the matching features are eliminated by Random Sample Consensus (RANSAC) algorithm. To create seamless image, alpha blending operation is applied. Experiments are conducted on UDISD (Unsupervised Deep Image Stitching Data set). The overall performance of the proposed stitching method is evaluated based on metrics such as PSNR, SSIM, RMSE, MSE and UIQI, and the proposed stitching algorithm yields good result with seamless stitched image.

[...] Read more.
Fingerprint Image Fusion: A Cutting-edge Perspective on Gender Classification via Rotational Invariant Features

By Shivanand Gornale Abhijit Patil Khang Wen Goh Sathish Kumar Kruthi R

DOI: https://doi.org/10.5815/ijigsp.2024.04.04, Pub. Date: 8 Aug. 2024

In this cutting-edge technological milieu, fingerprints have become an alternative expression for the biometrics system. A fingerprint is one of the perceptible biometric modals which is predominantly utilized in almost all the security, and real-life applications. Fingerprints have many inherent rotational features that are mostly utilized for person recognition besides these features can also be utilized for the person gender classification. Thus, the proposed work is a novel algorithm which identifies the gender of an individual based on the fingerprint. The image fusion and feature level fusion technique are deliberated over the fingerprints with rotational invariant features. Experiments were carried on four state-of-the-art datasets and realized promising results by outperforming earlier outcomes.

[...] Read more.
50Hz Power Line Interference Removal from an Electrocardiogram Signal Using a VME-DWT-Based Frequency Extraction and Filtering Approach

By Pavan G. Malghan Malaya Kumar Hota

DOI: https://doi.org/10.5815/ijigsp.2024.04.05, Pub. Date: 8 Aug. 2024

Removing undesirable artifacts in electrocardiogram signals is essential for biological signal processing as the signal gets distorted and makes appropriate investigation challenging. A primary source of distortion affecting recordings is the 50Hz power line interference. To get a high-quality recording, we used a filtering method based on an efficient decomposition technique known as variational mode extraction. This approach is similar to the variational mode decomposition methodology but with a few alterations in mathematical computation. First, it extracts the noise efficiently in a specific frequency band. Then, we apply the discrete wavelet transform to the signal, employing soft thresholding. As a result, it eliminates the extra noise and filters the electrocardiogram signal. We evaluated the efficacy of our proposed method using an arrhythmia database. Furthermore, we compared recent decomposition methods on six random signals using signal-to-noise ratios, mean square errors, correlation coefficients, and other signal features. Our method also efficiently eliminates varying amplitude of powerline noise and finally outperforms decomposition strategies regarding noise reduction and processing complexity across all signal parameters.

[...] Read more.
Fetal Brain Planes Classification Using Deep Ensemble Transfer Learning from U-Net Segmented Fetal Neurosonography Images

By Md. Nazmul Hasan A. B. M. Aowlad Hossain

DOI: https://doi.org/10.5815/ijigsp.2024.04.06, Pub. Date: 8 Aug. 2024

Fetal neurosonography is potentially used to examine the fetal brain by scanning the trans-thalamic (TT), trans-cerebellum (TC), and trans-ventricular (TV) planes. Cross-sectional analysis of these planes is useful to assess the brain anatomy, development, and abnormality for intervention and treatment plans even at the postnatal stage. To minimize the errors and processing time involved in the traditional manual subjective approach, the automatic classification of fetal brain planes is crucial. In this study, a deep learning-based method for automatically categorizing fetal brain planes from ultrasound images is proposed and evaluated. Firstly, the brain region has been segmented from the fetal brain ultrasound images using U-Net to prepare an efficient data set for the classifier model. Then, an ensemble convolutional neural network (CNN) model including well-known Inception V3, ResNet50-V2, and DenseNet-201 models with max voting is designed to classify the segmented brain planes. 2019 fetal brain ultrasound images from a widely used publicly accessible experts-annotated dataset are used to evaluate the performance of the proposed framework. The obtained results analysis shows that using the segmented images as input improves the performance of the classifier from its raw form. The gradient class activation mapping (Grad-CAM) based inspection shows noteworthy localization capability of the last convolution layer. The ensemble model has also outperformed its individual model’s performance. The suggested categorization framework is satisfactory compared to related recent works, with a testing accuracy of 97.68%. The proposed framework for fetal brain plane classification is expected to be useful for clinical applications.

[...] Read more.
Computerized Acute Myeloid Leukemia Classification Using Hybrid Dilated DenseSqueeze Network from Peripheral B Stain Analysis

By Krishna Prasad Palli Edara Sreenivasa Reddy Chandra Sekharaiah K.

DOI: https://doi.org/10.5815/ijigsp.2024.04.07, Pub. Date: 8 Aug. 2024

In medical diagnosis, Artificial Intelligence (AI) has offered significant revolution, especially for cancers. Acute Myeloid Leukemia (AML) is a deadly blood cancer caused by the rapid growth of abnormal White Blood Cells (WBCs) in humans. Although AML classification is a popular area of research, existing detection methods utilize manual examination of microscopic blood samples, which includes high complexity and tedious. Therefore, this work presented a computerized deep learning model-based AML classification from peripheral blood stain images, which helps in earlier AML diagnosis. The processing steps followed in AML classification are Image Pre-processing, Localization of RoI (Region of Interest), Fusion-based Feature Extraction and Classification. First, the input image is pre-processed, which includes noise filtering, image resizing, and colour conversion. The noise in the image is filtered using normalized Gaussian filtering (NGF). Next, the image is resized into a standard size, and the RGB image is converted into CMYK colour space. Then, the RoI is identified using the Image Moment Localization (IML) technique. Next, the valuable multi-level dense features are extracted using DenseSqueeze Network, and multi-scale features are extracted using Dilated Convolution Spatial Pyramid Pooling (Dilated CSPP). Both these extracted features are fused using the element-wise summation. Finally, the Softmax classifier is used in the last layer to classify the classes of AML and the loss in the network is optimized using the Improved Artificial Fish Swarm (Improved AFS) algorithm. The proposed work results in 99% of accuracy, 98.5% of precision and 98.9% of F-score by using the AML-Cytomorphology LMU dataset.

[...] Read more.
Autism Spectrum Disorder Screening on Home Videos Using Deep Learning

By Anjali Singh Abha Rawat Mitali Laroia Seeja K. R.

DOI: https://doi.org/10.5815/ijigsp.2024.04.08, Pub. Date: 8 Aug. 2024

Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder characterized by difficulty in social interactions, communication, and repetitive behaviors. Protocols like ADOS (Autism Diagnostic Observation Schedule) and ADI-R (Autism Diagnostic Interview Revised) are used by experts to assess the subject’s behavior which is time-consuming. Over the decade, researchers have studied the application of various Machine Learning techniques for ASD diagnosis through facial feature analysis, eye movement tracking, questionnaire analysis, functional magnetic resonance imaging (fMRI) analysis, etc. However, these techniques are not helpful for the parent or guardian of the child to perform an initial screening. This research proposes a novel deep learning model to diagnose ASD using general videos of the subject performing some tasks with the parent/guardian. Since there is no publicly available dataset on ASD videos, a dataset is created by collecting the videos of autistic children performing some activities with parents/guardians from YouTube from different demographic locations. These videos are then converted to skeletal key points to extract the child's engagement and social interaction in a given task. The proposed CNN-LSTM model is trained on 80% of the collected videos and then tested on the remaining 20%. The experiment results on various combinations of pre-trained CNN models and LSTM/BiLSTM show that the proposed model can be used as an initial autism screening tool. Among the different combinations, the MobileNet and Bi-LSTM combo achieved the best test accuracy of 84.95% with 89% precision, recall and F1-score.

[...] Read more.