Data Augmentation Strategy Research Articles

DeepRice: A deep learning and deep feature based classification of Rice leaf disease subtypes

Rice stands as a crucial staple food globally, with its enduring sustainability hinging on the prompt detection of rice leaf diseases. Hence, efficiently detecting diseases when they have already occurred holds paramount importance for solving the cost of manual visual identification and chemical testing. In the recent past, the identification of leaf pathologies in crops predominantly relies on manual methods using specialized equipment, which proves to be time-consuming and inefficient. This study offers a remedy by harnessing Deep Learning (DL) and transfer learning techniques to accurately identify and classify rice leaf diseases. A comprehensive dataset comprising 5932 self-generated images of rice leaves was assembled along with the benchmark datasets, categorized into 9 classes irrespective of the extent of disease spread across the leaves. These classes encompass diverse states including healthy leaves, mild and severe blight, mild and severe tungro, mild and severe blast, as well as mild and severe brown spot. Following meticulous manual labelling and dataset segmentation, which was validated by horticulture experts, data augmentation strategies were implemented to amplify the number of images. The datasets were subjected to evaluation using the proposed tailored Convolutional Neural Networks models. Their performance are scrutinized in conjunction with alternative transfer learning approaches like VGG16, Xception, ResNet50, DenseNet121, Inception ResnetV2, and Inception V3. The effectiveness of the proposed custom VGG16 model was gauged by its capacity to generalize to unseen images, yielding an exceptional accuracy of 99.94%, surpassing the benchmarks set by existing state-of-the-art models. Further, the layer wise feature extraction is also visualized as the interpretable AI.

Multi-branch stacking remote sensing image target detection based on YOLOv5

Optical remote sensing is crucial in land management, maritime safety, and rescue operations. Currently, high resolution target detection faces the problems including feature loss, false detection, and limited network robustness. To tackle the aforementioned issues, this study introduces a novel MBS-NET model, which is built upon the YOLOv5 structure. The proposed model presents a multi-branch stacking module structure to precisely capture deep target feature information. By introducing a dual-channel attention mechanism module (EGCA) at the model's neck, the proposed method ensures discriminative feature acquisition in crowded objects and vast spatial scenes. Prior to network training, introduce an improved data augmentation strategy to accommodate the multi-scale and directional variations in remote sensing objectives for model detection. Experimental results on the large-scale public DIOR dataset demonstrate that the MBS-NET model introduced in this paper displays exceptional performance and remarkable interpretability in remote sensing scenarios at a large scale. MBS-NET model outperforms YOLOv5 and YOLOv7 models by increasing the accuracy by 5% and 2% respectively. In addition, the recall rate and F1 Score index of MBS-NET model is superior to those of other methods, resulting in significant improvement of detection accuracy and robustness in large scenes.

DA-VEGAN: Differentiably Augmenting VAE-GAN for microstructure reconstruction from extremely small data sets

Microstructure reconstruction is an important and emerging field of research and an essential foundation to improving inverse computational materials engineering (ICME). Much of the recent progress in the field is made based on generative adversarial networks (GANs). Although excellent results have been achieved throughout a variety of materials, challenges remain regarding the interpretability of the model’s latent space as well as the applicability to extremely small data sets. The present work addresses these issues by introducing DA-VEGAN, a model with two central innovations. First, a β-variational autoencoder is incorporated into a hybrid GAN architecture that allows to penalize strong nonlinearities in the latent space by an additional parameter, β. Secondly, a custom differentiable data augmentation scheme is developed specifically for this architecture. The differentiability allows the model to learn from extremely small data sets without mode collapse or deteriorated sample quality. An extensive validation on a variety of structures demonstrates the potential of the method and future directions of investigation are discussed.

Enhancing coal-gangue object detection using GAN-based data augmentation strategy with dual attention mechanism

Coal separation based on computer vision has attracted substantial attention in recent years. However, developing reliable object detection models relies on large-scale annotated dataset, which in industrial practice is time-consuming and labor-intensive to obtain. In this paper, we propose a novel data augmentation model called dual attention deep convolutional generative adversarial network (DADCGAN) to expand dataset scale and improve object detection. For the first time, the proposed DADCGAN, which adopts DCGAN as its foundation architecture, introduces efficient channel attention and external attention mechanisms to capture essential feature information from the channel and spatial dimensions of images, respectively. Moreover, spectral normalization and two time-scale update rule strategies are incorporated to stabilize the training process. The implementation of our proposed data augmentation strategy includes two steps. First, traditional pixel transformation is used to expand an original small dataset. Then, our GAN-based data augmentation is executed to further expand the dataset by generating synthetic images. Experimental results show that our DADCGAN model achieves the lowest FID value, decreasing the FID by 21.30–71.96 % compared to other baseline GAN models, showcasing its ability to produce more realistic coal-gangue images. Finally, the data augmentation strategies are applied to the YOLOv4 model, enhancing the mAP by 9.26 %, highlighting its significance in enhancing coal-gangue object detection. These results have important implications for the development and implementation of computer vision-based technologies, enabling the realization of cleaner and more efficient coal separation methods.

DeepSSM: A blueprint for image-to-shape deep learning models

Statistical shape modeling (SSM) characterizes anatomical variations in a population of shapes generated from medical images. Statistical analysis of shapes requires consistent shape representation across samples in shape cohort. Establishing this representation entails a processing pipeline that includes anatomy segmentation, image re-sampling, shape-based registration, and non-linear, iterative optimization. These shape representations are then used to extract low-dimensional shape descriptors that are anatomically relevant to facilitate subsequent statistical analyses in different applications. However, the current process of obtaining these shape descriptors from imaging data relies on human and computational resources, requiring domain expertise for segmenting anatomies of interest. Moreover, this same taxing pipeline needs to be repeated to infer shape descriptors for new image data using a pre-trained/existing shape model. Here, we propose DeepSSM, a deep learning-based framework for learning the functional mapping from images to low-dimensional shape descriptors and their associated shape representations, thereby inferring statistical representation of anatomy directly from 3D images. Once trained using an existing shape model, DeepSSM circumvents the heavy and manual pre-processing and segmentation required by classical models and significantly improves the computational time, making it a viable solution for fully end-to-end shape modeling applications. In addition, we introduce a model-based data-augmentation strategy to address data scarcity, a typical scenario in shape modeling applications. Finally, this paper presents and analyzes two different architectural variants of DeepSSM with different loss functions using three medical datasets and their downstream clinical application. Experiments showcase that DeepSSM performs comparably or better to the state-of-the-art SSM both quantitatively and on application-driven downstream tasks. Therefore, DeepSSM aims to provide a comprehensive blueprint for deep learning-based image-to-shape models.

Effective image tampering localization with multi-scale ConvNeXt feature fusion

With the widespread use of powerful image editing tools, image tampering becomes easy and realistic. Existing image forensic methods still face challenges of low generalization performance and robustness. In this letter, we propose an effective image tampering localization scheme based on ConvNeXt encoder and multi-scale Feature Fusion (ConvNeXtFF). Stacked ConvNeXt blocks are utilized as an encoder to capture hierarchical multi-scale features, which are then fused in decoder for locating tampered pixels accurately. Combined loss function and effective data augmentation strategies are adopted to further improve the model performance. Extensive experimental results show that both localization accuracy and robustness of the ConvNeXtFF scheme outperform other state-of-the-art ones. The source code is available at https://github.com/multimediaFor/ConvNeXtFF.

Instance attack: an explanation-based vulnerability analysis framework against DNNs for malware detection

Deep neural networks (DNNs) are increasingly being used in malware detection and their robustness has been widely discussed. Conventionally, the development of an adversarial example generation scheme for DNNs involves either detailed knowledge concerning the model (i.e., gradient-based methods) or a substantial quantity of data for training a surrogate model. However, under many real-world circumstances, neither of these resources is necessarily available. Our work introduces the concept of the instance-based attack, which is both interpretable and suitable for deployment in a black-box environment. In our approach, a specific binary instance and a malware classifier are utilized as input. By incorporating data augmentation strategies, sufficient data are generated to train a relatively simple and interpretable model. Our methodology involves providing explanations for the detection model, which entails displaying the weights assigned to different components of the specific binary. Through the analysis of these explanations, we discover that the data subsections have a significant impact on the identification of malware. In this study, a novel function preserving transformation algorithm designed specifically for data subsections is introduced. Our approach involves leveraging binary diversification techniques to neutralize the effects of the most heavily-weighted section, thus generating effective adversarial examples. Our algorithm can fool the DNNs in certain cases with a success rate of almost 100%. Instance attack exhibits superior performance compared to the state-of-the-art approach. Notably, our technique can be implemented in a black-box environment and the results can be verified utilizing domain knowledge. The model can help to improve the robustness of malware detectors.

BiMGCL: rumor detection via bi-directional multi-level graph contrastive learning

The rapid development of large language models has significantly reduced the cost of producing rumors, which brings a tremendous challenge to the authenticity of content on social media. Therefore, it has become crucially important to identify and detect rumors. Existing deep learning methods usually require a large amount of labeled data, which leads to poor robustness in dealing with different types of rumor events. In addition, they neglect to fully utilize the structural information of rumors, resulting in a need to improve their identification and detection performance. In this article, we propose a new rumor detection framework based on bi-directional multi-level graph contrastive learning, BiMGCL, which models each rumor propagation structure as bi-directional graphs and performs self-supervised contrastive learning based on node-level and graph-level instances. In particular, BiMGCL models the structure of each rumor event with fine-grained bidirectional graphs that effectively consider the bi-directional structural characteristics of rumor propagation and dispersion. Moreover, BiMGCL designs three types of interpretable bi-directional graph data augmentation strategies and adopts both node-level and graph-level contrastive learning to capture the propagation characteristics of rumor events. Experimental results on real datasets demonstrate that our proposed BiMGCL achieves superior detection performance compared against the state-of-the-art rumor detection methods.

SFAMNet: A scene flow attention-based micro-expression network

Tremendous progress has been made in facial Micro-Expression (ME) spotting and recognition; however, most works have focused on either spotting or recognition tasks on the 2D videos. Until recently, the estimation of the 3D motion field (a.k.a scene flow) for the ME has only become possible after the release of the multi-modal ME dataset. In this paper, we propose the first Scene Flow Attention-based Micro-expression Network, namely SFAMNet. It takes the scene flow computed using the RGB-D flow algorithm as the input and predicts the spotting confidence score and emotion labels. Specifically, SFAMNet is an attention-based end-to-end multi-stream multi-task network devised to spot and recognize the ME. Besides that, we present a data augmentation strategy to alleviate the small sample size problem during network learning. Extensive experiments are performed on three tasks: (i) ME spotting; (ii) ME recognition; and (iii) ME analysis on the multi-modal CAS(ME)3 dataset. Empirical results indicate that depth is vital in capturing the ME information and the effectiveness of the proposed approach. Our source code is publicly available at https://github.com/genbing99/SFAMNet.

Development of efficient techniques for ASR System for Speech Detection and Recognization system using Gaussian Mixture Model- Universal Background Model

Some practical uses of ASR have been implemented, including the transcription of meetings and the usage of smart speakers. It is the process by which speech waves are transformed into text that allows computers to interpret and act upon human speech. Scalable strategies for developing ASR systems in languages where no voice transcriptions or pronunciation dictionaries exist are the primary focus of this work. We first show that the necessity for voice transcription into the target language can be greatly reduced through cross-lingual acoustic model transfer when phonemic pronunciation lexicons exist in the new language. Afterwards, we investigate three approaches to dealing with languages that lack a pronunciation lexicon. Secondly, we have a look at the efficiency of graphemic acoustic model transfer, which makes it easy to build pronunciation dictionaries. Thesis problems can be solved, in part, by investigating optimization strategies for training on huge corpora (such as GA+HMM and DE+HMM). In the training phase of acoustic modelling, the suggested method is applied to traditional methods. Read speech and HMI voice experiments indicated that while each data augmentation strategy alone did not always increase recognition performance, using all three techniques together did. Power normalised cepstral coefficient (PNCC) features are tweaked somewhat in this work to enhance verification accuracy. To increase speaker verification accuracy, we suggest employing multiple “Gaussian Mixture Model-Universal Background Model (GMM-UBM) and SVM classifiers”. Importantly, pitch shift data augmentation and multi-task training reduced bias by more than 18% absolute compared to the baseline system for read speech, and applying all three data augmentation techniques during fine tuning reduced bias by more than 7% for HMI speech, while increasing recognition accuracy of both native and non-native Dutch speech.

Deep Reinforcement Learning Based Resource Allocation and Trajectory Planning in Integrated Sensing and Communications UAV Network

In this paper, multi-UAVs serve as mobile aerial ISAC platforms to sense and communicate with on-ground target users. To optimize the communication and sensing performance, we formulate a joint user association, UAV trajectory planning and power allocation problem to maximize the minimum weighted spectral efficiency among UAVs. This paper exploits the centralized and the decentralized deep reinforcement learning (DRL) solutions to solve the sequential decision-making problem. On one hand, we first introduce the centralized soft actor-critic (SAC) algorithm. Then, we explore the equivalent transformation of the optimization objective based on symmetric group, propose the random and the adaptive data augmentation schemes to design the replay memory buffer of SAC, and accordingly propose SAC algorithms assisted by data augmentation to tackle the transformed problem. On the other hand, the multi-agent soft actor-critic (MASAC), a decentralized solution, is also introduced to solve this sequential decision-making problem. The experiment results reveal the effectiveness of the centralized and the decentralized solutions in considered scenarios. Specifically, the SAC assisted by the adaptive scheme significantly outperforms other centralized solutions in the training speed and the weighted spectral efficiency. Meanwhile, the decentralized MASAC algorithm behaves best in the early training speed.

Exposing low-quality deepfake videos of Social Network Service using Spatial Restored Detection Framework

The increasing abuse of facial manipulation methods, such as FaceSwap, Deepfakes etc., seriously threatens the authenticity of digital images/videos on the Internet. Therefore, it is of great importance to identify the facial videos to confirm the contents and avoid fake news or rumours. Many researchers have paid great attention to the detection of deepfakes and put forward a number of deep-learning-based detection models.The existing approaches mostly face the performance degradation in detecting low-quality(LQ) videos, i.e. heavily compressed or low-resolution videos through some SNS (Social Network Service), resulting in the limitation in real applications. To address this issue, in this paper, a novel Spatial Restore Detection Framework(SRDF) is proposed for improving the detection performance for LQ videos by restoring spatial features. We designed a feature extraction-enhancement block and a mapping block inspired by super-resolution methods, to restore and enhance texture features. An attention module was introduced to guide the texture features restoration and enhancement stage attending to different local areas and restoring the texture features. Besides, an improved isolated loss was put forward to prevent the expansion of a single area concerned. Moreover, we adopted a regional data augmentation strategy to prompt feature restore and enhancement in the region attended. Extensive experiments conducted on two deepfake datasets have validated the superiority of the proposed method compared to the state-of-the-art, especially in the scenarios of detecting low-quality deepfake videos.

Joint Optimization of Class-Specific Training- and Test-Time Data Augmentation in Segmentation.

This paper presents an effective and general data augmentation framework for medical image segmentation. We adopt a computationally efficient and data-efficient gradient-based meta-learning scheme to explicitly align the distribution of training and validation data which is used as a proxy for unseen test data. We improve the current data augmentation strategies with two core designs. First, we learn class-specific training-time data augmentation (TRA) effectively increasing the heterogeneity within the training subsets and tackling the class imbalance common in segmentation. Second, we jointly optimize TRA and test-time data augmentation (TEA), which are closely connected as both aim to align the training and test data distribution but were so far considered separately in previous works. We demonstrate the effectiveness of our method on four medical image segmentation tasks across different scenarios with two state-of-the-art segmentation models, DeepMedic and nnU-Net. Extensive experimentation shows that the proposed data augmentation framework can significantly and consistently improve the segmentation performance when compared to existing solutions. Code is publicly available1.

Photovoltaic cell defect classification based on integration of residual-inception network and spatial pyramid pooling in electroluminescence images

Electroluminescence (EL) imaging provides high spatial resolution and better identifies micro-defects for inspection of photovoltaic (PV) modules. However, the analysis of EL images could be typically a challenging process due to complex defect patterns and inhomogeneous background structure. In this study, a deep convolutional neural network (CNN) model using residual connections and spatial pyramid pooling (SPP) is proposed for the efficient classification of PV cell defects. The proposed CNN model is built on the Inception-v3 network. In this way, feature maps in inception modules are shared to reuse in deeper layers and the representation ability of features is enriched with the pooling process of the SPP in different sizes. Due to the imbalanced class distribution, offline data augmentation strategies are applied and network performance is further improved. The proposed method is evaluated on a publicly available dataset of 8 classes, of which 7 classes are defective and one class is defect-free images. In the comparative evaluation, while other approaches give accuracy values between 76.49% and 89.17%, this value is increased to 93.59% with the proposed method. The experimental results show that the proposed method exhibits more accurate and robust classification performance compared with other model combinations and CNN models.

OpenMix+: Revisiting Data Augmentation for Open Set Recognition

Open set recognition requires models to recognize samples of known classes learned in the training set while reject unknowns not learned. Compared with the structural risk minimization theory for closed-set problems, structural risk in open set tasks remains rarely explored. In this paper, we point out that balancing between structural risk and open space risk is crucial for open set recognition, and re-formalize it as open set structural risk. This brings a new view towards the general relationship between closed set recognition and open set recognition against the common intuition, which argues that a good closed set classifier always benefits for open set recognition. Specifically, we theoretically and experimentally show that recent mix-based data augmentation methods are aggressive closed set regularization methods, which reduce structural risk at cost of sacrificing open space risk. Besides, we show that existing negative data augmentation designed for open space risk reduction also ignore the trade-off problem between structural risk and open space risk, which limits their performance. We propose an efficient negative data augmentation strategy named self-mix and a corresponding method named OpenMix. OpenMix generates high-quality negative samples by mixing samples themselves, which can take care of both risks simultaneously. When combining OpenMix with conservative closed set regularization methods to form OpenMix+, models can achieve lower open set structural risk. Extensive experiments validate the superiority of OpenMix and OpenMix+ in terms of both effectiveness and universality.

Advanced Human Activity Recognition through Data Augmentation and Feature Concatenation of Micro-Doppler Signatures

Developing accurate classification models for radar-based Human Activity Recognition (HAR), capable of solving real-world problems, depends heavily on the amount of available data. In this paper, we propose a simple, effective, and generalizable data augmentation strategy along with preprocessing for micro-Doppler signatures to enhance recognition performance. By leveraging the decomposition properties of the Discrete Wavelet Transform (DWT), new samples are generated with distinct characteristics that do not overlap with those of the original samples. The micro-Doppler signatures are projected onto the DWT space for the decomposition process using the Haar wavelet. The returned decomposition components are used in different configurations to generate new data. Three new samples are obtained from a single spectrogram, which increases the amount of training data without creating duplicates. Next, the augmented samples are processed using the Sobel filter. This step allows each sample to be expanded into three representations, including the gradient in the x-direction (Dx), y-direction (Dy), and both x- and y-directions (Dxy). These representations are used as input for training a three-input convolutional neural network-long short-term memory support vector machine (CNN-LSTM-SVM) model. We have assessed the feasibility of our solution by evaluating it on three datasets containing micro-Doppler signatures of human activities, including Frequency Modulated Continuous Wave (FMCW) 77 GHz, FMCW 24 GHz, and Impulse Radio Ultra-Wide Band (IR-UWB) 10 GHz datasets. Several experiments have been carried out to evaluate the model's performance with the inclusion of additional samples. The model was trained from scratch only on the augmented samples and tested on the original samples. Our augmentation approach has been thoroughly evaluated using various metrics, including accuracy, precision, recall, and F1-score. The results demonstrate a substantial improvement in the recognition rate and effectively alleviate the overfitting effect. Accuracies of 96.47%, 94.27%, and 98.18% are obtained for the FMCW 77 GHz, FMCW 24 GHz, and IR- UWB 10 GHz datasets, respectively. The findings of the study demonstrate the utility of DWT to enrich micro-Doppler training samples to improve HAR performance. Furthermore, the processing step was found to be efficient in enhancing the classification accuracy, achieving 96.78%, 96.32%, and 100% for the FMCW 77 GHz, FMCW 24 GHz, and IR-UWB 10 GHz datasets, respectively.

DALib: A Curated Repository of Libraries for Data Augmentation in Computer Vision.

Data augmentation is a fundamental technique in machine learning that plays a crucial role in expanding the size of training datasets. By applying various transformations or modifications to existing data, data augmentation enhances the generalization and robustness of machine learning models. In recent years, the development of several libraries has simplified the utilization of diverse data augmentation strategies across different tasks. This paper focuses on the exploration of the most widely adopted libraries specifically designed for data augmentation in computer vision tasks. Here, we aim to provide a comprehensive survey of publicly available data augmentation libraries, facilitating practitioners to navigate these resources effectively. Through a curated taxonomy, we present an organized classification of the different approaches employed by these libraries, along with accompanying application examples. By examining the techniques of each library, practitioners can make informed decisions in selecting the most suitable augmentation techniques for their computer vision projects. To ensure the accessibility of this valuable information, a dedicated public website named DALib has been created. This website serves as a centralized repository where the taxonomy, methods, and examples associated with the surveyed data augmentation libraries can be explored. By offering this comprehensive resource, we aim to empower practitioners and contribute to the advancement of computer vision research and applications through effective utilization of data augmentation techniques.

Missing well-log reconstruction using a sequence self-attention deep-learning framework

Well logging is a critical tool for reservoir evaluation and fluid identification. However, due to borehole conditions, instrument failure, economic constraints, etc., some types of well logs are occasionally missing or unreliable. Existing logging curve reconstruction methods based on empirical formulas and fully connected deep neural networks (FCDNN) can only consider point-to-point mapping relationships. Recurrently structured neural networks can consider a multipoint correlation, but it is difficult to compute in parallel. To take into account the correlation between log sequences and achieve computational parallelism, we develop a novel deep-learning framework for missing well-log reconstruction based on state-of-the-art transformer architecture. The missing well-log transformer (MWLT) uses a self-attention mechanism instead of a circular recursive structure to model the global dependencies of the inputs and outputs. To use different usage requirements, we design the MWLT in three scales: small, base, and large, by adjusting the parameters in the network. A total of 8609 samples from 209 wells in the Sichuan Basin, China, are used for training and validation, and two additional blind wells are used for testing. The data augmentation strategy with random starting points is implemented to increase the robustness of the model. The results show that our proposed MWLT achieves a significant improvement in accuracy over the conventional Gardner’s equation and data-driven approaches such as FCDNN and bidirectional long short-term memory, on the validation data set and blind test wells. The MWLT-large and MWLT-base have lower prediction errors than MWLT-small but require more training time. Two wells in the Songliao Basin, China, are used to evaluate the cross-regional generalized performance of our method. The generalizability test results demonstrate that density logs reconstructed by MWLT remain the best match to the observed data compared with other methods. The parallelizable MWLT automatically learns the global dependence of the parameters of the subsurface reservoir, enabling an efficient missing well-log reconstruction performance.

DeepHipp: accurate segmentation of hippocampus using 3D dense-block based on attention mechanism

BackgroundThe hippocampus is a key area of the brain responsible for learning, memory, and other abilities. Accurately segmenting the hippocampus and precisely calculating the volume of the hippocampus is of great significance for predicting Alzheimer’s disease and amnesia. Most of the segmentation algorithms currently involved are based on templates, such as the more popular FreeSufer.MethodsThis study proposes Deephipp, a deep learning network based on a 3D dense block using an attention mechanism for accurate segmentation of the hippocampus. DeepHipp is based on the following novelties: (i) DeepHipp adopts powerful data augmentation schemes to enhance the segmentation ability. (ii) DeepHipp is designed to incorporate 3D dense-block to capture multiple-scale features of the hippocampus. (iii) DeepHipp creatively uses the attention mechanism in the field of hippocampal image segmentation, extracting useful hippocampus information in a massive feature map, and improving the accuracy and sensitivity of the model.ConclusionsWe describe the illustrative results and show extensive qualitative and quantitative comparisons with other methods. Our achievement demonstrates that the accuracy of DeepHipp can reach 83.63%, which is superior to most existing methods in terms of accuracy and efficiency of hippocampus segmentation. It is noticeable that deep learning can potentially lead to an effective segmentation of medical images.

MixNet-LD: An Automated Classification System for Multiple Lung Diseases Using Modified MixNet Model.

The lungs are critical components of the respiratory system because they allow for the exchange of oxygen and carbon dioxide within our bodies. However, a variety of conditions can affect the lungs, resulting in serious health consequences. Lung disease treatment aims to control its severity, which is usually irrevocable. The fundamental objective of this endeavor is to build a consistent and automated approach for establishing the intensity of lung illness. This paper describes MixNet-LD, a unique automated approach aimed at identifying and categorizing the severity of lung illnesses using an upgraded pre-trained MixNet model. One of the first steps in developing the MixNet-LD system was to build a pre-processing strategy that uses Grad-Cam to decrease noise, highlight irregularities, and eventually improve the classification performance of lung illnesses. Data augmentation strategies were used to rectify the dataset's unbalanced distribution of classes and prevent overfitting. Furthermore, dense blocks were used to improve classification outcomes across the four severity categories of lung disorders. In practice, the MixNet-LD model achieves cutting-edge performance while maintaining model size and manageable complexity. The proposed approach was tested using a variety of datasets gathered from credible internet sources as well as a novel private dataset known as Pak-Lungs. A pre-trained model was used on the dataset to obtain important characteristics from lung disease images. The pictures were then categorized into categories such as normal, COVID-19, pneumonia, tuberculosis, and lung cancer using a linear layer of the SVM classifier with a linear activation function. The MixNet-LD system underwent testing in four distinct tests and achieved a remarkable accuracy of 98.5% on the difficult lung disease dataset. The acquired findings and comparisons demonstrate the MixNet-LD system's improved performance and learning capabilities. These findings show that the proposed approach may effectively increase the accuracy of classification models in medicinal image investigations. This research helps to develop new strategies for effective medical image processing in clinical settings.