Convolutional Neural Networks

978-93-7143-144-6 363 Chapter -23 Convolutional Neural Networks Mr. VIGNESHWAR RANGINI Department: IT Data Warehouse Solution Architect Capgemini, USA Email id: rangini [email protected] Abstract Convolutional Neural Networks (CNNs) have revolutionized the field of computer vision and deep learning, establishing themselves as the dominant architecture for processing grid-structured data. This chapter provides a comprehensive examination of CNN architecture, fundamental principles, and applications across diverse domains. We explore the mathematical foundations underlying convolutional operations, pooling mechanisms, and hierarchical feature learning. The chapter discusses architectural innovations from AlexNet to modern transformer-hybrid models, examining key developments that have shaped contemporary deep learning. We analyze CNN applications in image classification, object detection, semantic segmentation, medical imaging, and natural language processing. Furthermore, we address critical challenges including computational efficiency, interpretability, adversarial robustness, and generalization capabilities. The chapter also presents emerging trends such as neural architecture search, lightweight CNN designs for edge computing, and integration with attention mechanisms. Through theoretical analysis and practical insights, this chapter serves as a comprehensive resource for researchers and practitioners seeking to understand and implement CNN-based solutions. Convolutional Neural Networks 364 Keywords: Convolutional Neural Networks, Deep Learning, Computer Vision, Feature Learning, Image Classification, Neural Network Architecture, Transfer Learning, Object Detection 1. Introduction Convolutional Neural Networks represent a specialized class of artificial neural networks designed to process data with grid-like topology, most notably images and video sequences. Since their inception, CNNs have fundamentally transformed computer vision, achieving unprecedented performance in tasks that were once considered exclusively within the domain of human perception. The architecture draws inspiration from the organization of the animal visual cortex, where individual neurons respond to stimuli in restricted regions of the visual field known as receptive fields. The modern resurgence of CNNs began with the landmark AlexNet architecture in 2012, which demonstrated that deep convolutional networks could achieve superior performance on large-scale image recognition tasks when trained with sufficient data and computational resources. This breakthrough catalyzed an explosion of research and development, leading to increasingly sophisticated architectures and expanding the application scope beyond computer vision to encompass natural language processing, speech recognition, drug discovery, and autonomous systems. The fundamental principle underlying CNNs is the exploitation of spatial and temporal structure in data through three key architectural ideas: local connectivity, parameter sharing, and equivariant representations. These principles enable CNNs to learn hierarchical feature representations efficiently, progressing from low-level features such as edges and textures to high-level semantic concepts. This chapter systematically examines the theoretical foundations, architectural components, training methodologies, and practical Mr. VIGNESHWAR RANGINI 365 applications of CNNs, providing readers with both conceptual understanding and implementation insights. 2. Theoretical Foundations of CNNs 2.1 Mathematical Formulation of Convolution The convolutional operation forms the cornerstone of CNN architecture. Mathematically, a discrete convolution in two dimensions is defined as the weighted sum of input values within a local neighborhood. For an input feature map X and a kernel (filter) K, the convolution operation produces an output feature map Y where each element is computed by sliding the kernel across the input and performing element-wise multiplication followed by summation. In formal notation, the output of a convolutional layer can be expressed as Y(i,j) = Σ Σ X(i+m, j+n) × K(m,n) + b, where the summation occurs over the kernel dimensions, and b represents a bias term. This operation preserves spatial relationships in the input while extracting local features through learned filters. The key advantage of convolution over fully connected operations is the dramatic reduction in parameters through weight sharing—the same kernel is applied across all spatial locations, making the network more efficient and reducing the risk of overfitting. The convolutional operation exhibits an important property called translation equivariance, meaning that if the input is translated, the output is translated by the same amount. This property is particularly valuable for vision tasks where objects may appear at different positions in an image. Additionally, multiple kernels can be applied in parallel to extract different features, with each kernel learning to detect specific patterns such as edges, corners, or textures at various orientations. Convolutional Neural Networks 366 2.2 Receptive Field and Hierarchical Feature Learning The receptive field refers to the region of the input space that affects a particular unit in the network. As we progress deeper into a CNN, the effective receptive field grows, allowing higher layers to integrate information from increasingly larger regions of the input. This hierarchical organization enables CNNs to build complex, compositional representations from simple primitives. Early layers with small receptive fields detect local features such as edges and color gradients, while deeper layers with larger receptive fields recognize more abstract patterns like object parts and complete objects. The growth of receptive field size can be controlled through several architectural choices. Stacking multiple convolutional layers with small kernels (e.g., 3×3) can achieve the same receptive field as a single layer with a larger kernel, but with fewer parameters and increased non-linearity due to additional activation functions between layers. Pooling operations also contribute to receptive field expansion while providing translation invariance and reducing computational burden by downsampling the feature maps. 2.3 Activation Functions and Non-linearity Non-linear activation functions are essential for enabling neural networks to learn complex, non-linear mappings between inputs and outputs. Without non-linearity, stacking multiple layers would be equivalent to a single linear transformation. The Rectified Linear Unit (ReLU) has become the dominant activation function in modern CNNs due to its computational efficiency and ability to mitigate the vanishing gradient problem. ReLU is defined as f(x) = max(0, x), effectively zeroing out negative values while preserving positive activations unchanged. Several variants of ReLU have been proposed to address its limitations, including the "dying ReLU" problem where neurons can Mr. VIGNESHWAR RANGINI 367 become permanently inactive during training. Leaky ReLU introduces a small slope for negative values, while Parametric ReLU (PReLU) makes this slope learnable. Exponential Linear Unit (ELU) and Scaled Exponential Linear Unit (SELU) provide smooth, nonzero gradients for negative inputs and have shown benefits in specific applications. More recently, Swish and GELU activation functions, which incorporate smooth, non-monotonic characteristics, have demonstrated improved performance in very deep networks. 3. Core Architectural Components 3.1 Convolutional Layers: Design and Variations Convolutional layers constitute the primary building blocks of CNNs, performing feature extraction through learnable filters. Several hyperparameters govern the behavior of convolutional layers: kernel size determines the spatial extent of filters, stride controls the step size when sliding filters across the input, and padding specifies whether and how to extend input boundaries. Common kernel sizes include 3×3, 5×5, and 7×7, with modern architectures favoring smaller kernels stacked in sequence to achieve greater depth with fewer parameters. Stride directly affects the spatial dimensions of the output feature maps—larger strides produce smaller outputs, providing computational efficiency at the cost of spatial resolution. Padding strategies, particularly "same" padding that preserves input dimensions, prevent excessive information loss at boundaries. The number of filters in a layer determines the number of output feature maps, with deeper layers typically employing more filters to capture increasing representational complexity. Standard practice involves progressively increasing filter counts while reducing spatial dimensions through pooling or strided convolutions. Convolutional Neural Networks 368 Advanced convolutional variants have emerged to address specific challenges. Dilated (atrous) convolutions introduce spacing between kernel elements, expanding the receptive field without increasing parameters. Depthwise separable convolutions, popularized by MobileNets, factorize standard convolutions into depthwise and pointwise operations, dramatically reducing computational cost. Grouped convolutions partition input channels into groups, performing separate convolutions on each group before concatenating results, enabling more efficient computation and encouraging diverse feature learning. 3.2 Pooling Operations and Dimensionality Reduction Pooling layers perform spatial downsampling, reducing feature map dimensions while retaining essential information. The two most common pooling operations are max pooling, which selects the maximum value within each pooling window, and average pooling, which computes the mean. Max pooling has gained wider adoption due to its ability to preserve the strongest activations, making it more robust to small translations and distortions in the input. Typical pooling windows are 2×2 with stride 2, halving both height and width dimensions. Pooling provides several crucial benefits: it introduces partial translation invariance, reducing sensitivity to exact feature positions; it decreases computational requirements by reducing feature map sizes; and it increases effective receptive field sizes in subsequent layers. However, pooling also has limitations—it discards spatial information that may be important for precise localization tasks. Recent architectures have experimented with alternatives such as strided convolutions for downsampling, or elimination of pooling entirely in favor of increased depth, demonstrating that pooling is not strictly necessary for effective feature learning. Mr. VIGNESHWAR RANGINI 369 3.3 Batch Normalization and Layer Normalization Batch Normalization (BN) has become a standard component in modern CNN architectures, addressing the internal covariate shift problem by normalizing layer inputs across mini-batches during training. By standardizing activations to have zero mean and unit variance, BN stabilizes the learning process, enables higher learning rates, reduces sensitivity to initialization, and provides a mild regularization effect. The normalization is followed by learnable scale and shift parameters, allowing the network to recover the original distribution if beneficial. Mathematically, for a mini-batch of activations, BN computes the mean and variance, normalizes the activations, and applies affine transformation: normalized_x = gamma × (x - mean) / sqrt(variance + epsilon) + beta, where gamma and beta are learned parameters. During inference, running statistics accumulated during training replace batch statistics to ensure consistent behavior. Alternative normalization schemes include Layer Normalization, which normalizes across channel dimensions rather than batch dimensions, and Group Normalization, which divides channels into groups for normalization, offering advantages when batch sizes are small. 3.4 Fully Connected Layers and Classification Heads Fully connected (dense) layers traditionally serve as the classification head in CNNs, mapping learned feature representations to output predictions. After feature extraction through convolutional and pooling layers, the spatial feature maps are typically flattened into a one-dimensional vector and passed through one or more fully connected layers. The final layer uses softmax activation for multi-class classification, producing a probability distribution over classes, or sigmoid activation for binary classification and multi-label scenarios. Convolutional Neural Networks 370 However, fully connected layers contain a large proportion of network parameters and lack translation equivariance, making them computationally expensive and potentially prone to overfitting. Modern architectures increasingly adopt Global Average Pooling (GAP) as an alternative, averaging each feature map to a single value before classification. GAP drastically reduces parameters, acts as a structural regularizer, and maintains spatial information throughout the network. Some architectures employ hybrid approaches, using a single fully connected layer after GAP for additional representational capacity while maintaining efficiency. 4. Evolution of CNN Architectures 4.1 AlexNet: The Deep Learning Revolution AlexNet, introduced by Krizhevsky, Sutskever, and Hinton in 2012, marked a watershed moment in computer vision and deep learning. Competing in the ImageNet Large Scale Visual Recognition Challenge (ILSVRC), AlexNet achieved a top-5 error rate of 15.3%, substantially outperforming the second-place entry and previous traditional computer vision approaches. The architecture consisted of eight learned layers: five convolutional layers followed by three fully connected layers, with ReLU activations, max pooling, and dropout regularization. Several key innovations contributed to AlexNet's success. The use of ReLU activation functions instead of traditional sigmoid or tanh enabled training of deeper networks by mitigating gradient vanishing. Data augmentation through random crops, horizontal flips, and color jittering artificially expanded the training dataset, improving generalization. Dropout, applied to fully connected layers, randomly deactivated neurons during training to prevent coadaptation and reduce overfitting. The training was distributed across two GPUs, a practical necessity that also influenced the architecture design. AlexNet's success catalyzed intense research Mr. VIGNESHWAR RANGINI 371 interest in deep learning and established CNNs as the dominant paradigm for visual recognition. 4.2 VGGNet: Depth and Simplicity VGGNet, developed by the Visual Geometry Group at Oxford University in 2014, demonstrated that network depth is a critical factor for performance. The architecture pursued simplicity and uniformity, using only 3×3 convolutional filters throughout the entire network while significantly increasing depth. VGG-16 and VGG-19 variants achieved 16 and 19 weight layers respectively, showing steady performance improvements with increased depth. The consistent use of small kernels allowed VGGNet to have a large effective receptive field while maintaining fewer parameters than equivalent architectures with larger kernels. VGGNet followed a simple and regular design pattern: stacks of convolutional layers with 3×3 filters, followed by max pooling layers with 2×2 windows. The number of filters doubled after each pooling operation, starting from 64 and reaching 512 in deeper layers. This systematic design philosophy made VGGNet easy to understand and implement, contributing to its widespread adoption. Despite its simplicity, VGGNet required substantial memory and computation due to the large number of parameters, particularly in the fully connected layers, spurring subsequent research into more parameter-efficient architectures. 4.3 ResNet: Residual Learning and Skip Connections Residual Networks (ResNet), introduced by He et al. in 2015, addressed the degradation problem where deeper networks paradoxically performed worse than shallower counterparts, not due to overfitting but difficulty in optimization. ResNet introduced skip connections (shortcut connections) that bypass one or more layers, enabling the network to learn residual functions with reference to layer inputs. Instead of learning the desired underlying mapping Convolutional Neural Networks 378 layers to preserve spatial information. U-Net introduced encoderdecoder architecture with skip connections, combining highresolution low-level features with semantic high-level features, achieving excellent performance particularly in medical image segmentation with limited training data. DeepLab series advanced segmentation through atrous convolutions controlling receptive field size, atrous spatial pyramid pooling capturing multi-scale context, and conditional random fields refining boundary predictions. Applications include autonomous driving requiring pixel-accurate understanding of road scenes; medical image analysis delineating organs and pathological regions; satellite image interpretation for land cover classification and urban planning; video editing enabling background replacement; and agricultural monitoring assessing crop health. Instance segmentation extends semantic segmentation by distinguishing individual object instances, essential for counting and tracking applications. 6.4 Medical Image Analysis and Diagnosis CNNs have revolutionized medical imaging analysis, achieving performance rivaling human experts in specific diagnostic tasks. Applications span radiology (detecting tumors, fractures, and abnormalities in X-rays, CT, and MRI scans), pathology (identifying cancerous cells in tissue samples), ophthalmology (screening for diabetic retinopathy and age-related macular degeneration), and dermatology (classifying skin lesions). The hierarchical feature learning capability of CNNs proves particularly valuable for medical images where relevant patterns span multiple scales and may be subtle or complex. Challenges specific to medical imaging include limited labeled data due to the cost and expertise required for annotation, class imbalance with rare pathological cases, and stringent requirements for interpretability and reliability in clinical deployment. Transfer Mr. VIGNESHWAR RANGINI 379 learning from natural images, data augmentation, synthetic data generation, and semi-supervised learning help address data scarcity. Attention mechanisms and gradient-based visualization techniques improve interpretability by highlighting diagnostically relevant regions. Integration of CNN systems into clinical workflows promises to enhance diagnostic accuracy, reduce reading time, enable screening at scale, and democratize access to expert-level analysis in resource-limited settings. 7. Current Challenges and Limitations 7.1 Interpretability and Explainability The black-box nature of deep CNNs poses significant challenges for understanding and trusting their decisions, particularly in highstakes applications such as medical diagnosis, autonomous vehicles, and criminal justice. Interpretability techniques aim to elucidate what features the network learns and how it reaches particular decisions. Visualization methods such as activation maximization, saliency maps, and Grad-CAM highlight input regions that most influence predictions. These techniques reveal that CNNs learn hierarchical representations but may rely on spurious correlations or textures rather than semantic content humans would consider relevant. Attention mechanisms provide inherent interpretability by learning to weight feature importance. Network dissection techniques probe individual neurons to understand their semantic meaning. Concept activation vectors identify human-interpretable concepts in learned representations. Despite these advances, fundamental tensions exist between model complexity necessary for high performance and interpretability—simpler models are more interpretable but less accurate, while complex models achieve better performance but remain opaque. Developing inherently interpretable architectures that maintain competitive performance remains an active research direction critical for trustworthy AI deployment. Convolutional Neural Networks 380 7.2 Adversarial Robustness and Security CNNs exhibit troubling vulnerability to adversarial examples— inputs with imperceptibly small perturbations that cause misclassification. These adversarial perturbations can be crafted through optimization methods that maximize prediction error while constraining perturbation magnitude. The existence of adversarial examples reveals that neural networks learn decision boundaries that differ fundamentally from human perception, potentially exploiting statistical artifacts in training data rather than robust semantic features. This vulnerability poses security risks in adversarial settings where malicious actors might deliberately craft inputs to evade detection or cause incorrect behavior. Adversarial training, incorporating adversarial examples during training, improves robustness but at the cost of decreased accuracy on clean data and increased computational requirements. Certified defenses provide formal guarantees of robustness within specified perturbation bounds, though often with substantial performance penalties. Input preprocessing, ensemble methods, and detection schemes offer additional protection but can be circumvented by adaptive attacks. The fundamental tension between accuracy and robustness, formalized through adversarial risk bounds, suggests that inherent limitations may prevent simultaneous optimization of both objectives. Developing truly robust models remains a critical challenge for deploying CNNs in security-critical applications. 7.3 Data Efficiency and Few-Shot Learning Modern CNNs typically require massive labeled datasets for training, limiting their applicability in domains where data collection or annotation is expensive, time-consuming, or restricted by privacy concerns. Humans demonstrate remarkable ability to learn from few examples, while CNNs generalize poorly when trained on small datasets, often overfitting to training data without learning transferable representations. This data inefficiency stems Mr. VIGNESHWAR RANGINI 381 partly from CNNs learning superficial statistical patterns rather than causal relationships, and from optimization algorithms requiring extensive sampling of the input distribution. Few-shot learning aims to enable models to adapt to new tasks with minimal examples, using techniques such as meta-learning (learning to learn), metric learning (learning similarity functions), and transfer learning (adapting pretrained models). Self-supervised learning extracts supervisory signals from unlabeled data through pretext tasks such as predicting image rotations, solving jigsaw puzzles, or contrastive learning, enabling models to learn useful representations without manual labels. Semi-supervised learning leverages both labeled and unlabeled data, using techniques like pseudo-labeling and consistency regularization. Despite progress, closing the gap between human and machine data efficiency remains a fundamental challenge requiring new learning paradigms beyond current deep learning approaches. 7.4 Computational Costs and Environmental Impact Training state-of-the-art CNNs requires substantial computational resources, with costs growing super-linearly with model size and dataset scale. Training large vision models can consume thousands of GPU-hours, corresponding to significant energy usage and carbon emissions. This computational burden limits research to well-funded institutions, raises environmental concerns, and creates barriers to entry for developing regions and smaller organizations. Inference costs also matter for deployed systems, particularly in resourceconstrained environments like mobile devices or edge computing scenarios where power consumption directly impacts battery life. Addressing these challenges requires multi-pronged approaches: developing more efficient architectures like MobileNets and EfficientNets that achieve comparable performance with fewer operations; model compression techniques including pruning, Convolutional Neural Networks 382 quantization, and knowledge distillation that reduce model size and computational requirements; neural architecture search to automatically discover efficient architectures; and specialized hardware accelerators optimized for deep learning workloads. Green AI initiatives advocate for reporting energy consumption and carbon footprint alongside accuracy metrics, encouraging development of efficient rather than merely accurate models. Balancing performance, efficiency, and environmental sustainability represents a crucial consideration for responsible AI development. 8. Emerging Trends and Future Directions 8.1 Vision Transformers and Attention Mechanisms Vision Transformers (ViTs) represent a paradigm shift in computer vision, applying transformer architectures originally developed for natural language processing directly to images. Unlike CNNs that process images through local convolutional operations, ViTs divide images into patches, embed them as sequences, and process them through self-attention layers that can capture long-range dependencies in a single operation. When trained on sufficiently large datasets, ViTs match or exceed CNN performance while offering more interpretable attention patterns and greater architectural flexibility. Hybrid architectures combining convolutional layers for low-level feature extraction with transformers for high-level reasoning show promising results, suggesting complementary strengths of local inductive bias from convolutions and global context modeling from attention. Swin Transformer introduces hierarchical representations and shifted windows for efficient attention computation, achieving state-of-the-art performance on various vision tasks. However, transformers require substantially more training data than CNNs due to reduced inductive bias, and their quadratic computational complexity with respect to sequence length poses scalability challenges. Ongoing research explores efficient attention Mr. VIGNESHWAR RANGINI 383 mechanisms, better training strategies for limited data regimes, and optimal integration of convolution and attention. 8.2 Neural Architecture Search and AutoML Neural Architecture Search (NAS) automates the design of neural network architectures through algorithmic search over architectural choices, potentially discovering novel designs that outperform handcrafted architectures. NAS methods define a search space of possible architectures, a search strategy to explore this space (such as reinforcement learning, evolutionary algorithms, or gradient-based optimization), and a performance estimation strategy to evaluate candidate architectures. Successful NAS-discovered architectures like EfficientNet and NAS-Net demonstrate that automated search can find architectures that achieve better accuracy-efficiency tradeoffs than manual design. However, NAS faces significant challenges: computational costs can be prohibitive, requiring thousands of GPU-days for single searches; searched architectures may not transfer well to different datasets or tasks; and the search space design itself requires expert knowledge, limiting true automation. One-shot NAS and weightsharing strategies dramatically reduce search costs by training a single supernet encompassing all candidate architectures. Differentiable NAS formulates architecture search as continuous optimization, enabling gradient-based search. Hardware-aware NAS considers deployment constraints like latency and energy consumption during search. As NAS methods mature, they promise to democratize architecture design and automatically adapt networks to specific applications and hardware platforms. 8.3 Self-Supervised and Contrastive Learning Self-supervised learning extracts supervisory signals from unlabeled data, enabling models to learn rich representations without manual annotation. Contrastive learning methods like SimCLR, MoCo, and Convolutional Neural Networks 384 BYOL train models to distinguish between different augmented views of the same image (positive pairs) and views from different images (negative pairs), learning representations that capture semantic content invariant to augmentation. These methods have achieved remarkable success, with self-supervised pretraining on unlabeled images rivaling or exceeding supervised pretraining on labeled datasets for downstream transfer learning. Masked image modeling, inspired by masked language modeling in NLP, randomly masks image patches and trains models to predict masked content from context. This approach, exemplified by MAE (Masked Autoencoders), demonstrates that simple pretext tasks can yield powerful representations. Self-supervised learning promises to unlock the value of vast quantities of unlabeled visual data available on the internet and in domain-specific applications. Future directions include developing pretext tasks better aligned with downstream objectives, combining multiple self-supervised signals, and extending self-supervised methods to video and multimodal data for learning richer temporal and cross-modal representations. 8.4 Multimodal Learning and Vision-Language Models Multimodal learning aims to build models that understand and generate content across multiple modalities, particularly vision and language. Vision-language models like CLIP learn joint representations of images and text through contrastive learning on large-scale image-caption datasets, enabling zero-shot image classification by computing similarity between images and textual class descriptions. These models demonstrate impressive generalization, transferring to novel visual concepts described in natural language without additional training. Extensions support tasks including image captioning, visual question answering, textto-image generation, and image-text retrieval. Mr. VIGNESHWAR RANGINI 385 Large-scale vision-language models trained on diverse internet data exhibit emergent capabilities such as visual reasoning, object composition understanding, and semantic alignment between modalities. Grounding language in visual perception promises more robust language understanding and common-sense reasoning. Challenges include efficiently scaling to multiple modalities beyond vision and language, handling modality-specific versus shared information, and developing unified architectures that seamlessly integrate different input types. Multimodal learning represents a step toward more general artificial intelligence systems that perceive and interact with the world through multiple complementary sensory channels, similar to human cognition. 9. Conclusion Convolutional Neural Networks have fundamentally transformed computer vision and established themselves as a cornerstone of modern artificial intelligence. From the pioneering AlexNet to contemporary architectures incorporating attention mechanisms and neural architecture search, CNNs have evolved through continuous innovation in architectural design, training methodologies, and application domains. The hierarchical feature learning paradigm underlying CNNs—progressing from simple edge detectors to complex semantic representations—has proven remarkably effective across diverse tasks including classification, detection, segmentation, and generation. Despite tremendous successes, significant challenges remain. Improving interpretability and explainability is essential for trustworthy deployment in critical applications. Enhancing adversarial robustness protects against security vulnerabilities. Increasing data efficiency through self-supervised learning and fewshot methods broadens applicability to data-scarce domains. Reducing computational requirements addresses environmental concerns and democratizes access to advanced AI capabilities. Convolutional Neural Networks 386 These challenges drive active research exploring novel architectures, training paradigms, and theoretical understanding of deep learning. Looking forward, the integration of CNNs with transformers, the automation of architecture design through NAS, and the expansion into multimodal learning represent exciting frontiers. The convergence of computer vision with natural language processing enables richer human-computer interaction and more general AI systems. As CNN technology matures and limitations are addressed, we anticipate increasingly sophisticated applications spanning healthcare, autonomous systems, scientific discovery, creative tools, and beyond. The journey from early neocognitrons to modern deep learning systems illustrates the power of biologically-inspired computation, and future advances promise to further extend the capabilities and impact of convolutional neural networks across all facets of society and technology. References 1. Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). ImageNet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems (pp. 1097-1105). 2. LeCun, Y., Bottou, L., Bengio, Y., & Haffner, P. (1998). Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11), 2278-2324. 3. Simonyan, K., & Zisserman, A. (2014). Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556. 4. He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (pp. 770-778). Mr. VIGNESHWAR RANGINI 387 5. Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., ... & Rabinovich, A. (2015). Going deeper with convolutions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (pp. 1-9). 6. Howard, A. G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., ... & Adam, H. (2017). MobileNets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861. 7. Ronneberger, O., Fischer, P., & Brox, T. (2015). U-Net: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention (pp. 234-241). Springer. 8. Long, J., Shelhamer, E., & Darrell, T. (2015). Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (pp. 3431-3440). 9. Redmon, J., Divvala, S., Girshick, R., & Farhadi, A. (2016). You only look once: Unified, real-time object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (pp. 779-788). 10. Ren, S., He, K., Girshick, R., & Sun, J. (2015). Faster RCNN: Towards real-time object detection with region proposal networks. In Advances in Neural Information Processing Systems (pp. 91-99). 11. Huang, G., Liu, Z., Van Der Maaten, L., & Weinberger, K. Q. (2017). Densely connected convolutional networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (pp. 4700-4708). 12. Ioffe, S., & Szegedy, C. (2015). Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International Conference on Machine Learning (pp. 448-456).