This study presents a unified low-parameter approach to multi-class classification of microorganisms (micrococci, diplococci, streptococci, and bacilli) based on automated machine learning. The method is designed to produce interpretable taxonomic descriptors through analysis of the external geometric characteristics of microorganisms, including cell shape, colony organization, and dynamic behavior in unfixed microscopic scenes. A key advantage of the proposed approach is its lightweight nature: the resulting models have significantly fewer parameters than deep learning-based alternatives, enabling fast inference even on standard CPU hardware. An annotated dataset containing images of four bacterial types obtained under conditions simulating real clinical trials has been developed and published to validate the method. The results (Precision = 0.910, Recall = 0.901, and F1-score = 0.905) confirm the effectiveness of the proposed method for biomedical diagnostic tasks, especially in settings with limited computational resources and a need for feature interpretability. Our approach demonstrates performance comparable to state-of-the-art methods while offering superior efficiency and lightweight design due to its significantly reduced number of parameters.
25 June 2025 in Pattern Recognition and Image Analysis
This paper focuses on advancing the field of small object detection within complex visual environments, leveraging the latest in deep learning technologies. We introduce a novel approach characterized by a dual-stream self-attention mechanism integrated within a multihead framework, and further refine detection accuracy through an innovative output reweighting technique. The core of our methodology, termed ADSAR (advanced dual-stream attention and reweighting), is designed to tackle the challenges posed by small objects that often overlap multiple tokens in feature maps—a common issue in conventional detection models. By dynamically adjusting the scale of attention across different heads, ADSAR allows for detailed feature capture at multiple granularities, significantly enhancing the model’s ability to detect and characterize small objects. The addition of a softmax-based reweighting function selectively emphasizes features crucial for object recognition, thereby suppressing irrelevant information and reducing noise. Our proposed model not only outperforms existing state-of-the-art solutions in accuracy but also demonstrates superior efficiency in processing and scalability. These advancements contribute to the theoretical understanding of attention mechanisms in deep neural networks and offer practical improvements in real-world applications where small object detection is critical.
1 December 2024 in Pattern Recognition and Image Analysis
This paper presents a study aimed at improving the detection quality of small-sized microorganisms under challenging microscopic conditions through the application of a lightweight combined image preprocessing model. We focused on the task of detecting diplococci in images obtained through dynamic sample microscopy. The proposed approach employs predefined filters for image preprocessing, combined with the prediction of their applicability. This enables the elimination of artifacts such as blurred boundaries and indistinct edges, which are often encountered when working with small-sized objects in nonfixed scenes. The application of the proposed model has significantly improved detection quality compared to traditional methods.
Timely identification and accurate delineation of ultra-early ischemic stroke lesions in non-contrast computed tomography (CT) scans of the human brain are of paramount importance for prompt medical intervention and improved patient outcomes. In this study, we propose a deep learning-driven methodology specifically designed for segmenting ultra-early ischemic regions, with a particular emphasis on both the ischemic core and the surrounding penumbra during the initial stages of stroke progression. We introduce a lightweight preprocessing model based on convolutional filtering techniques, which enhances image clarity while preserving the structural integrity of medical scans, a critical factor when detecting subtle signs of ultra-early ischemic strokes. Unlike conventional preprocessing methods that directly modify the image and may introduce artifacts or distortions, our approach ensures the absence of neural network-induced artifacts, which is especially crucial for accurate diagnosis and segmentation of ultra-early ischemic lesions. The model employs predefined differentiable filters with trainable parameters, allowing for artifact-free and precision-enhanced image refinement tailored to the challenges of ultra-early stroke detection. In addition, we incorporated into the combined preprocessing pipeline a newly proposed trainable linear combination of pretrained image filters, a concept first introduced in this study. For model training and evaluation, we utilize a publicly available dataset of acute ischemic stroke cases, focusing on the subset relevant to ultra-early stroke manifestations, which contains annotated non-contrast CT brain scans from 112 patients. The proposed model demonstrates high segmentation accuracy for ultra-early ischemic regions, surpassing existing methodologies across key performance metrics. The results have been rigorously validated on test subsets from the dataset, confirming the effectiveness of our approach in supporting the early-stage diagnosis and treatment planning for ultra-early ischemic strokes.
25 June 2025 in Pattern Recognition and Image Analysis
This study explores the incorporation of specialized self-attention mechanisms into deep learning architectures, with a particular emphasis on segmenting human iris and pupil regions in infrared images. In this work, we present some modified versions of nonlocal blocks designed to enhance self-attentive properties while addressing the distinct characteristics of infrared imaging data. By applying these customized enhancements, we have observed significant improvements in the performance metrics of the underlying deep neural network model, with segmentation results showing substantial improvement (increasing from 0.945 to 0.983 in mIoU and from 0.951 to 0.988 in mDice) after thorough testing on a representative infrared image dataset. This advancement paves the way for various applications in infrared image analysis, offering new opportunities for research and innovation.