A Comprehensive Review Of Image Processing Techniques: Classical Methods, Deep Learning, Applications, And Recent Advances

Figure 1. Evolution of Image Processing from Classical Algorithms to Foundation Models

Figure 1 should visually represent the technological progression of image processing from pixel-level computational operations to increasingly sophisticated AI-based visual understanding. Unlike the existing Figure 1, which primarily presents a general processing workflow, the revised figure should emphasize the historical and technological evolution of the field. The current manuscript already illustrates the conventional progression from image acquisition through preprocessing, enhancement, segmentation, feature extraction, classification, and real-world applications.

2. Materials and Methods

2.1 Review Design

The present study adopted a structured narrative literature review with thematic and comparative analysis to examine the development of digital image processing from conventional computational techniques to contemporary artificial intelligence (AI)-based approaches. The review was structured to cover the major stages of image processing, including image representation, enhancement, restoration, segmentation, feature extraction, compression, deep learning, Transformer-based methods, generative models, explainable artificial intelligence (XAI), and foundation models. Particular attention was given to the relationship between methodological characteristics, computational requirements, application domains, and practical deployment. The organization of the review follows the broad progression identified in the source paper, from classical image-processing operators toward CNNs, Transformers, generative models, and foundation-model approaches.

The review is narrative rather than a formal systematic review, because the available source manuscript does not report a registered protocol, complete database-search record, PRISMA flow diagram, or quantitative meta-analysis. Therefore, no numerical pooled effect size or statistical meta-analysis was performed.

2.2 Literature Sources

Relevant literature was considered from major scholarly information sources and publisher platforms, including IEEE Xplore, ScienceDirect, SpringerLink, PubMed, ACM Digital Library, Google Scholar, and relevant journals published by major academic publishers. The literature base was supplemented by authoritative books and conference proceedings where they provided foundational or technically important contributions.

The principal search concepts included “digital image processing,” “image enhancement,” “image restoration,” “image segmentation,” “feature extraction,” “CNN image processing,” “Vision Transformer,” “deep learning image analysis,” “explainable AI image processing,” “foundation models computer vision,” and “image processing agriculture.” The existing paper provides a substantial reference base covering image processing fundamental, deep learning, medical imaging, remote sensing, agriculture, explainable AI, transfer learning, and foundation models.

2.3 Inclusion and Exclusion Criteria

Literature was considered relevant when it addressed one or more major aspects of digital image processing or computer vision. The inclusion criteria comprised studies that:

• described image-processing techniques or computational image analysis;
• presented computer vision or pattern-recognition methodologies;
• investigated CNN, transfer learning, Transformer, or related AI approaches;
• examined practical image-processing applications;
• introduced important methodological developments;
• provided methodological, comparative, or application-related information; and
• represented peer-reviewed articles, authoritative books, or recognized conference publications.

The exclusion criteria comprised duplicate publications, sources unrelated to image processing, non-technical opinion material, and publications that did not provide sufficient methodological information for meaningful classification or comparison. Duplicate versions of the same study were also excluded from thematic consideration.

2.4 Thematic Classification

The selected literature was organized into major methodological themes to facilitate comparison. These categories were derived from the technical structure of the reviewed paper and its reference literature.

This classification enabled the comparison of traditional handcrafted approaches with automatically learned visual representations and newer general-purpose vision architectures.

2.5 Comparative Evaluation Framework

The reviewed approaches were compared according to accuracy, precision, recall, F1-score, computational complexity, dataset requirements, interpretability, processing speed, memory requirements, application domain, and deployment suitability. These measures were treated as comparative criteria reported or discussed in the literature rather than as results generated from a new experimental dataset.

For classification studies, the principal evaluation measures are defined as:

Accuracy

Precision

Recall

F1-Score

where TP represents true positives, TN true negatives, FP false positives, and FN false negatives. These measures provide complementary information regarding classification performance and are commonly used when comparing image-analysis approaches.

For image-quality assessment, the source paper also identifies Mean Square Error (MSE), Peak Signal-to-Noise Ratio (PSNR), and Structural Similarity Index Measure (SSIM) as important evaluation measures (Gonzalez & Woods, 2024; Wang et al., 2004).

Methodological note: Because this is a review study, the comparative tables and graphs in the Results and Discussion section should clearly distinguish reported values extracted from published studies from qualitative synthesis or illustrative information. This avoids presenting hypothetical values as original experimental findings.

3. Results and Discussion

The thematic analysis of the reviewed literature demonstrates that digital image processing has developed from relatively simple pixel-level operations to complex AI-driven systems capable of learning visual representations, performing semantic segmentation, generating images, and supporting multimodal visual reasoning. The reviewed approaches can broadly be classified into image representation, enhancement and restoration, segmentation, feature extraction, compression, deep learning, Transformer-based processing, explainable AI, generative AI, and foundation-model approaches. The analysis also indicates that no single technique is universally suitable for all applications; rather, the selection of an image-processing method depends on image characteristics, computational resources, dataset availability, required accuracy, interpretability, and deployment environment.

3.1 Fundamental Image Representation

Digital images are composed of pixels arranged according to spatial dimensions such as width and height. Each pixel represents the visual information at a particular spatial location. The amount of information represented by an image depends on factors including spatial resolution, bit depth, and the number of color channels. Higher spatial resolution can preserve finer visual details but generally increases storage and computational requirements. Similarly, greater bit depth provides a larger range of intensity values and can preserve more subtle variations in image information (Gonzalez & Woods, 2024). The existing paper describes a digital image as a two-dimensional array of pixels and emphasizes the importance of resolution and color representation in image analysis.

Different color representations are useful for different image-processing tasks. RGB represents images through red, green, and blue channels and is widely used in cameras and displays. HSV separates hue, saturation, and value and can be advantageous for color-based segmentation. YCbCr separates luminance from chrominance and is widely associated with image and video compression. Grayscale represents an image through a single intensity channel, reducing computational and memory requirements.

Table 1. Comparison of Image Representation Models

Thus, image representation is not merely a preliminary technical choice; it influences the effectiveness and computational requirements of subsequent enhancement, segmentation, feature extraction, and classification operations.

3.2 Image Enhancement and Restoration

Image enhancement improves the visual quality or visibility of important image characteristics, whereas image restoration attempts to recover information degraded by noise, blur, or other imaging processes. Common enhancement approaches include Histogram Equalization (HE), Contrast Limited Adaptive Histogram Equalization (CLAHE), contrast stretching, and gamma correction.

Histogram Equalization redistributes intensity values over the available dynamic range and can improve the contrast of low-contrast images. CLAHE extends this principle by operating on local regions and limiting excessive contrast amplification. Contrast stretching expands a restricted intensity range, while gamma correction applies a nonlinear transformation to adjust image brightness (Gonzalez & Woods, 2024). The existing Figure 2 demonstrates the progression from an original image through low contrast to an enhanced image and includes corresponding intensity histograms.

Image restoration addresses degradation using mathematical or statistical models. Gaussian noise can originate from electronic sensor and transmission effects, salt-and-pepper noise appears as isolated dark and bright pixels, and speckle noise is commonly associated with imaging systems such as radar and ultrasound. Wiener filtering is useful for restoration under suitable noise assumptions, median filtering is particularly useful for impulse noise, and adaptive filters can adjust their behavior according to local image characteristics.

Figure 2. Before and After Image Enhancement

The figure should contain the corresponding histograms to demonstrate the redistribution of pixel intensities.

3.3 Image Segmentation

Segmentation divides an image into meaningful regions and is fundamental to object identification and analysis. Conventional approaches include global thresholding, Otsu thresholding, region growing, watershed segmentation, and K-Means clustering. These methods differ in the assumptions they make about intensity, color, connectivity, or region similarity. Otsu's method automatically determines a threshold based on between-class variance, while K-Means groups pixels according to similarity.

Deep learning introduced more adaptive segmentation approaches, including methods reviewed specifically for agricultural image segmentation (Lei et al., 2024). U-Net, for example, uses an encoder-decoder architecture to combine contextual information with spatial localization and has become an important architecture for biomedical segmentation (Ronneberger et al., 2015). More recently, foundation-model approaches have expanded segmentation beyond individual task-specific models. Segment Anything (SAM) introduced a promptable segmentation framework designed to generalize across a wide range of image segmentation tasks (Kirillov et al., 2023).

Consequently, segmentation has evolved from manually designed thresholds and region rules toward learned and increasingly general-purpose visual representations.

Figure 3. Comparative Image Segmentation Workflow

This figure should visually compare thresholding/Otsu/K-Means with U-Net and SAM rather than presenting them as equivalent algorithms.

3.4 Feature Extraction

Feature extraction converts visual information into numerical representations that can be used for classification, recognition, matching, or analysis. Traditional feature descriptors include SIFT, SURF, ORB, and HOG, while image characteristics such as edges, shape, and texture can also serve as important descriptors. SIFT provides scale- and rotation-invariant local features, SURF emphasizes computational efficiency, ORB provides a comparatively lightweight alternative, and HOG represents local gradient and edge orientations.

Deep learning changed feature extraction by allowing features to be learned automatically from training data. CNNs progressively learn low-level, intermediate, and high-level representations, reducing the need for manually designed descriptors. Transformer architectures further extend representation learning by modelling contextual relationships between image regions.

Table 2. Traditional and Deep Feature Extraction Approaches

Traditional descriptors therefore remain relevant where computational resources or training data are limited, while learned representations are particularly useful for complex visual patterns.

3.5 Image Compression

Image compression reduces storage requirements and transmission bandwidth. It can be classified into lossless and lossy compression. Lossless methods preserve the original information, whereas lossy methods remove information considered less important for visual representation.

JPEG is a widely used lossy compression standard associated with the Discrete Cosine Transform (DCT), while JPEG2000 uses wavelet-based processing. PNG provides lossless compression and is particularly useful when preservation of image information or transparency is required.

Compression efficiency can be expressed as:

Compression Ratio=Original Data SizeCompressed Data Size

Higher compression can reduce storage and transmission requirements, but excessive compression can introduce artifacts and reduce image fidelity. Therefore, image compression involves a practical balance between compression ratio, storage efficiency, transmission requirements, and image quality.

3.6 Deep Learning-Based Image Processing

Deep learning has substantially changed image processing by enabling automatic feature learning from large image datasets. CNNs use convolutional filters to extract spatial features, activation functions such as ReLU to introduce nonlinear transformations, pooling to reduce spatial dimensions, and fully connected or classification layers to produce predictions. The existing Figure 4 illustrates this general CNN workflow.

ResNet introduced residual connections that allow information to pass through shortcut pathways and facilitate the training of deeper networks (He et al., 2016). Efficient architectures subsequently focused on balancing predictive performance with computational requirements. EfficientNet introduced compound scaling across network depth, width, and input resolution (Tan & Le, 2019), while MobileNet architectures were designed specifically for resource-constrained environments. MobileNetV3 incorporated hardware-aware architecture search and optimization for mobile deployment (Howard et al., 2019).

Transfer learning is particularly relevant when task-specific datasets are limited. A pretrained model can provide reusable visual representations that can subsequently be adapted to a target application. The reviewed literature identifies transfer learning and optimized CNN training as important approaches for improving image-processing performance while reducing training requirements (Ankatwar, 2025; Yadav & Gajanan, 2026).

3.7 Vision Transformers and Explainable AI

Vision Transformers (ViTs) introduced a different approach to image representation by dividing an image into patches and processing those patches as a sequence. Self-attention enables relationships between image regions to be modelled, providing a mechanism for capturing broader contextual information (Dosovitskiy et al., 2021). Swin Transformer further developed hierarchical Transformer representations and has become an important architecture for visual recognition and dense prediction (Liu et al., 2022).

The increasing complexity of deep models has simultaneously increased interest in Explainable AI (XAI). Techniques such as Grad-CAM, saliency maps, feature visualization, and SHAP can assist researchers and users in understanding model predictions (Samek et al., 2021). Grad-CAM uses gradients associated with a target class to generate class-discriminative localization maps, allowing important image regions influencing a CNN prediction to be visualized (Selvaraju et al., 2017). XAI is especially relevant in sensitive domains such as medical diagnosis and agricultural disease detection, where understanding the visual evidence supporting a prediction can complement conventional performance metrics. Recent plant-disease XAI literature also emphasizes the importance of evaluating explanation methods and their validation practices (Ismail et al., 2026).

Figure 4. Explainable AI Workflow

The figure 4 illustrates the workflow of explainable AI, showing how an input image is processed by a deep learning model, followed by prediction, gradient analysis, generation of visual explanations such as Grad-CAM heatmaps, and human interpretation of the model’s decision.

3.8 Generative AI and Foundation Models

Generative AI has expanded image processing beyond analysis toward image generation and transformation. GANs employ generator and discriminator networks and can be applied to synthetic-image generation, augmentation, and image-to-image translation. Diffusion models use a progressive noise-removal process to generate images and have become an important generative modelling paradigm (Ho et al., 2020).

Foundation models represent another major transition. Rather than developing an independent model for every individual task, foundation models are pretrained on large-scale data and can potentially be adapted to multiple downstream applications. The foundation-model literature identifies both opportunities and challenges associated with this approach, including generalization, computational requirements, and responsible deployment (Bommasani et al., 2022). SAM provides an important computer-vision example through promptable segmentation (Kirillov et al., 2023).

3.9 Applications of Image Processing

The reviewed literature demonstrates broad applicability across scientific and industrial domains.

In medical imaging, image enhancement, segmentation, and classification support MRI, CT, and X-ray analysis (Litjens et al., 2017). In remote sensing, image processing supports land-use mapping, environmental monitoring, disaster assessment, and infrastructure analysis.

In agriculture, image processing supports crop monitoring, disease detection, pest identification, and precision farming. CNNs, transfer learning, multispectral imaging, XAI, and deep-learning-based segmentation methods (Lei et al., 2024) are increasingly being investigated for crop-health assessment, including cotton disease detection (Ankatwar & Dhawale, 2024, 2025, 2026; Gajanan & Preetham, 2026).

Other important applications include autonomous vehicles, where image processing assists in detecting pedestrians, vehicles, lanes, and traffic signs; security, where visual recognition supports biometric authentication and surveillance; and industrial inspection, where computer vision identifies manufacturing defects and quality problems.

3.10 Comparative Analysis

The overall literature indicates a transition from computationally simple, explicitly designed algorithms toward data-driven models with increasing representational capacity. However, the relationship between accuracy, computational cost, data requirements, explainability, and deployment must be considered when selecting a method.

Table 3. Comparative Analysis of Major Image Processing Approaches

*Qualitative synthesis of the reviewed literature, not original experimental measurements.

The comparison shows that classical methods generally require fewer computational resources and can be easier to interpret. Their effectiveness, however, may depend strongly on image quality, illumination, noise, and background complexity. Deep learning methods can learn more complex representations but normally require greater computational resources and suitable training data. The current paper similarly identifies conventional methods as valuable for preprocessing and resource-constrained environments while highlighting the increasing importance of deep learning, transfer learning, and optimized architectures.

3.11 Performance Visualization and Graphs

The existing manuscript contains a performance graph comparing thresholding, Otsu, K-Means, region growing, watershed, and U-Net using accuracy, precision, recall, and F1-score. However, the manuscript explicitly identifies the numerical values in this graph as illustrative and states that they should be replaced with quantitative results obtained from the reviewed studies.

Figure 5. Qualitative Computational Complexity of Image Processing Techniques

The figure presents a qualitative comparison of computational complexity across classical image-processing methods, feature extraction techniques, CNNs, Transformers, and foundation models, ranging from low to very high computational demand.

Figure 6. Evolution of Image Processing Technologies

The figure illustrates the progression of image processing from classical techniques and machine learning to CNNs, transfer learning, Vision Transformers, generative AI, and foundation models, highlighting the increasing capabilities and broader applications of modern visual-processing systems.

Figure 7. Major Application Domains of Image Processing

The figure 7 illustrates the major application areas of image processing, including healthcare, agriculture, remote sensing, industrial inspection, security and surveillance, autonomous transportation, multimedia, environmental monitoring, and biometrics.

3.12 Challenges

Several challenges remain despite the rapid development of image-processing technologies. Deep learning systems frequently depend on sufficiently large and representative labelled datasets. Dataset imbalance, annotation cost, limited diversity, and domain shift can influence model generalization. These issues become particularly important when models are transferred from laboratory or benchmark datasets to real-world environments.

Computational requirements are another challenge. Large CNNs, Transformers, and foundation models may require substantial GPU memory and processing capacity, making lightweight CNNs and AI edge devices important considerations for practical deployment (Sun et al., 2024). Such requirements can restrict deployment on smartphones, embedded devices, agricultural sensors, and other edge platforms.

Interpretability also remains important. High predictive performance does not automatically explain why a model produces a particular prediction. Consequently, XAI methods are increasingly incorporated into visual analysis systems. Privacy and security are particularly relevant for medical and surveillance applications, while unreliable network connectivity can limit the practicality of cloud-dependent systems.

3.13 Future Research Directions

The literature indicates several important research directions, including lightweight CNNs, EfficientNet and MobileNet architectures, edge AI, explainable AI, federated learning, multispectral image processing, multimodal vision, vision-language models (Zhang et al., 2024), foundation models, generative data augmentation, diffusion models, real-time image processing, and privacy-preserving computer vision. The existing review similarly identifies lightweight architectures, edge computing, federated learning, foundation models, and multimodal systems as emerging areas for future development.

Future systems are therefore likely to move toward hybrid architectures that combine classical preprocessing with efficient deep learning and Transformer-based representations. In agriculture, for example, image enhancement and segmentation can be combined with lightweight classification models and XAI for deployment on edge devices. In healthcare, segmentation and classification can be integrated with explainability mechanisms to provide visual evidence alongside predictions. Vision-language and other multimodal foundation-model approaches may further reduce the need to construct independent models for every visual task (Zhang et al., 2024), although their computational and reliability requirements require continued investigation.

Figure 8. Future Research Framework for Intelligent Image Processing