MobileNet V2 Research Articles

Segmentation and classification of white blood SMEAR images using modified CNN architecture

The classification and recognition of leukocytes or WBCs in blood smear images presents a key role in the corresponding diagnosis of specific diseases, such as leukemia, tumor, hematological disorders, etc. The computerized framework for automated segmentation & classification of WBCs nucleus contributes an important role for the recognition of WBCs related disorders. Therefore, this work emphasizes WBCs nucleus segmentation using modified U-Net architecture and the segmented WBCs nucleus are further classified into their subcategory i.e., basophil, eosinophil, neutrophil, monocyte and lymphocyte. The classification and nucleus characterization task has been performed using VGGNet and MobileNet V2 architecture. Initially, collected instances are passed to the preprocessing phase for image rescaling and normalization. The rescaled and normalized instances are passed to the U-Net model for nucleus segmentation. Extracted nucleus are forwarded to the classification phase for their class identifications. Furthermore, the functioning of the intended design will be compared with other modern methods. By the end of this study a successful model classifying various nucleus morphologies such as Basophil, Eosinophil, Lymphocyte, Monocyte and Neutrophil was obtained where overall test accuracy achieved was 97.0% for VGGNet classifier and 94.0% for MobileNet V2 classifier.

Impact of Pre-trained Weights on MobileNet Architectures for Animal Classification

This study aims to address the problem of improving animal image classification accuracy using different versions of MobileNet. Accurate animal classification plays a vital role in biodiversity protection, environmental monitoring, and agriculture. The research is significant because existing studies focus on specific models and datasets, leaving a gap in the comparative performance analysis of MobileNet versions. To address this issue, MobileNet V1, V2, and V3 models were utilized, both with and without ImageNet pre-trained weights. The models were trained on a dataset composed of 30,179 images from two sources, covering 13 animal categories. The experiment involved training the models over 10 epochs using a standard configuration of the TensorFlow framework, with accuracy serving as the primary evaluation metric.The results showed that MobileNet V3Large, with pre-trained weights, achieved the highest accuracy (97.43%), outperforming V1 and V2. Using pre-trained weights consistently enhanced performance, as models without pre-training exhibited lower accuracy and slower convergence. This study contributes by providing a comprehensive comparison of MobileNet versions in animal classification tasks, demonstrating the importance of pre-training and model architecture optimization for achieving high accuracy in image classification.

Sparse attention with residual pyramidal depthwise separable convolutional based malware detection with optimization mechanism

Recent developments indicate that malware programs present a significant risk in the security and privacy of cloud systems. Existing research in malware detection encounters numerous significant challenges due to the constantly changing and advanced characteristics of malware. Malware detection systems frequently experience high rates of false positives and false negatives, where legitimate applications are incorrectly identified as malware or actual malware remains undetected, which results in operational inefficiencies. Traditional signature-based approaches struggle in recognizing new or modified malware. Additionally, sophisticated malware types, such as file less malware, ransom ware, and rootkits pose detection challenges as they integrate deeply into systems or alter their behaviour to evade detection. These challenges highlight the urgent need for ongoing advancements in this field. Existing methods of malware detection that rely on signatures have been found to be both inefficient and slow in the context of cloud environments. Also, Existing studies have focused on detecting malware by analysing input API calls. But, these models have encountered challenges such as limited accuracy and difficulties in effectively classifying malware types. In contrast, Deep Learning (DL) have shown success by analysing malware behaviour through API calls, which yields encouraging results. Additionally, the data produced by API calls necessitates more computational resources for training. To address these challenges, a new deep learning-based malware detection approach utilizes 2D grayscale images derived from API calls, along with an effective tuning strategy has been proposed. Initially, data are collected from cloud malware dataset. Then, API calls are converted into 2D gray scale images in order to construct gray scale image dataset. After getting the gray scale image, pre-processing is performed to reduce high level noise and to enhance the quality of image by weighted mean filter and anisotropic filter, which helps to improve the performance in classification. Next, these images are then passed into the feature extraction stage to extract sufficient features with an effective integrated densely connected squeeze MobileNet v2 (Ef-DeSMob2), which reduces the dimensionality issue and increase the computational complexity. Then, the collected features are passed into the classification phase to detect normal and malware classes from the samples using sparse attention with residual pyramidal depth wise separable convolutional neural networks (SA:ResPyDSC), which focus to enhance the security and reliability of the model. Finally, the hyper parameters in the classifier model like weights, bias are properly fine-tuned by utilizing hybrid white shark beluga optimization algorithm (Hy-WBeOp). The experimental findings illustrate that the proposed method attains accuracy of 98.06%, precision of 97.99%, recall of 97.05%, f1-score of 96.08% and error metrics like MSE AT 0.08, RMSE of 0.27 and MAE of 0.21, which shows that the proposed method helps to classify the malware classes accurately with less error rates. The proposed approach outperforms with the existing techniques because of its great efficiency. Overall, this approach establish a strong malware detection system classification and enhance the reliability and effectiveness of protection against malicious attacks.

A Novel Deep Learning Framework Enhanced by Hybrid Optimization Using Dung Beetle and Fick’s Law for Superior Pneumonia Detection

Pneumonia is an inflammation of lung tissue caused by various infectious microorganisms and noninfectious factors. It affects people of all ages, but vulnerable age groups are more susceptible. Imaging techniques, such as chest X-rays (CXRs), are crucial in early detection and prompt action. CXRs for this condition are characterized by radiopaque appearances or sometimes a consolidation in the affected part of the lung caused by inflammatory secretions that replace the air in the infected alveoli. Accurate early detection of pneumonia is essential to avoid its potentially fatal consequences, particularly in children and the elderly. This paper proposes an enhanced framework based on convolutional neural network (CNN) architecture, specifically utilizing a transfer-learning-based architecture (MobileNet V1), which has outperformed recent models. The proposed framework is improved using a hybrid method combining the operation of two optimization algorithms: the dung beetle optimizer (DBO), which enhances exploration by mimicking dung beetles’ navigational strategies, and Fick’s law algorithm (FLA), which improves exploitation by guiding solutions toward optimal areas. This hybrid optimization effectively balances exploration and exploitation, significantly enhancing model performance. The model was trained on 7750 chest X-ray images. The framework can distinguish between healthy and pneumonia, achieving an accuracy of 98.19 ± 0.94% and a sensitivity of 98 ± 0.99%. The results are promising, indicating that this new framework could be used for the early detection of pneumonia with a low cost and high accuracy, especially in remote areas that lack expertise in radiology, thus reducing the mortality rate caused by pneumonia.

Stereo disparity estimation based on Siamese network

Abstract This paper proposes an approach for estimating stereo disparity based on a Siamese network. The existing Top-k disparity regression strategy may overlook the true disparity at object edges. To address this issue, an additional maximum disparity estimation value is introduced to optimize the Top-k strategy, creating the Top-k+ strategy. Furthermore, this research improves the MobileNetv2 (MV2) block by introducing an attention mechanism in the frequency domain, resulting in a more efficient FMV2 block that extracts high-frequency information such as textures and edges. Considering the excellent performance of the gated recurrent unit (GRU) in disparity optimization, the algorithm adopts an iterative optimization method based on GRU. Comparative studies pertaining to the Scene Flow, KITTI 2015, and ETH3D benchmark datasets show outstanding results in disparity estimation, with an end-point error of 0.58 pixels on the Scene Flow dataset, a D1-all error of 1.67% on the KITTI 2015 dataset, and an average disparity error of 0.22 pixels on the ETH3D dataset. Compared to other GRU-based iterative optimization algorithms, the proposed method not only exhibits significant performance advantages but also demonstrates a lightweight design.

Research on the Method of Imperfect Wheat Grain Recognition Utilizing Hyperspectral Imaging Technology

As the primary grain crop in China, wheat holds a significant position in the country’s agricultural production, circulation, consumption, and various other aspects. However, the presence of imperfect grains has greatly impacted wheat quality and, subsequently, food security. In order to detect perfect wheat grains and six types of imperfect grains, a method for the fast and non-destructive identification of imperfect wheat grains using hyperspectral images was proposed. The main contents and results are as follows: (1) We collected wheat grain hyperspectral data. Seven types of wheat grain samples, each containing 300 grains, were prepared to construct a hyperspectral imaging system for imperfect wheat grains, and visible near-infrared hyperspectral data from 2100 wheat grains were collected. The Savitzky–Golay algorithm was used to analyze the hyperspectral images of wheat grains, selecting 261 dimensional effective hyperspectral datapoints within the range of 420.61–980.43 nm. (2) The Successive Projections Algorithm was used to reduce the dimensions of the 261 dimensional hyperspectral datapoints, selecting 33 dimensional hyperspectral datapoints. Principal Component Analysis was used to extract the optimal spectral wavelengths, specifically selecting hyperspectral images at 647.57 nm, 591.78 nm, and 568.36 nm to establish the dataset. (3) Particle Swarm Optimization was used to optimize the Support Vector Machines model, Convolutional Neural Network model, and MobileNet V2 model, which were established to recognize seven types of wheat grains. The comprehensive recognition rates were 93.71%, 95.14%, and 97.71%, respectively. The results indicate that a larger model with more parameters may not necessarily yield better performance. The research shows that the MobileNet V2 network model exhibits superior recognition efficiency, and the integration of hyperspectral image technology with the classification model can accurately identify imperfect wheat grains.

Deep Learning Models for Dental Conditions Classification Using Intraoral Images

This paper presents the digitalization of dentistry medical records to support the dentist in the patient examination process. A dentist uses manual input to fill out the evaluation form by drawing and labeling each patient’s tooth condition based on their observations. Consequently, it takes too long to finish only one examination. For time efficiency, using AI-based digitalization technology can be a promising solution. To address the problem, we made and compared several classification models to recognize human dental conditions to help doctors analyze patient teeth. We apply the YOLOv5, MobileNet V2, and IONet (proposed CNN model) as deep learning models to recognize the five common human dental conditions: normal, filling, caries, gangrene radix, and impaction. We tested the ability of YOLO classification as an object detection model and compared it with classification models. We used a dataset of 3.708 intraoral dental images generated by various augmentation methods from 1.767 original images. We collected and annotated the dataset with the help of dentists. Furthermore, the dataset is divided into three parts: 90% of the total dataset is used as training and validation data, then divided again into 80% training data and 20% validation data. 10% of the total dataset will be used as testing data to compare classification performance. Based on our experiments, YOLOv5, as an object detection model, can classify dental conditions in humans better than the classification model. YOLOv5 produces an 82% accuracy testing value and performs better than the classification model. MobileNet V2 and IONet only get 80% and 70% testing accuracy. Although statistically, there is not much of a difference between the test accuracy values for YOLOv5 and MobileNet v2, the speed in classifying dental objects using YOLOv5 is more efficient, considering that YOLOv5 is an object detection model. There are still challenges with the deep learning technique used in this research, but these can be addressed in further development. A more complex model and the enlargement of more data, ensuring it is varied and balanced, can be used to address the limitations.

Blind assistance system for appliance control and public transport safety using CNN, MobileNet V2 and Yolo V8

The purpose of this research is to address the challenges faced by visually impaired individuals, particularly in handling household appliances independently. With approximately 285 million visually impaired individuals worldwide, technological solutions are crucial to enhancing their accessibility and independence. This paper introduces a Smart Assistance System designed to empower visually impaired individuals to interact with household appliances in real-time without assistance. In this study, three Convolutional Neural Network (CNN) algorithms are compared to develop the system. The evaluation metrics include accuracy, precision, recall, F1 score, and hamming loss on validation images. The performance comparison reveals that the custom architecture CNN, MobileNetv2, and YOLO models achieve F1 scores of 0.43, 0.63, and 0.24, respectively. To enhance object detection and classification, the paper suggests implementing bounding box buttons categorization using YOLOv8, which demonstrates superior performance with a 95% classification accuracy on testing images of home appliance buttons. They face similar difficult while in public and accessing public property. Expanding upon the proposed system’s capabilities, the paper introduces the concept of panic button detection and activation in a bus environment tailored for blind individuals. This system relies on various factors such as the number of people onboard, heart rate monitoring, and the detection of distress signals or SOS sounds emitted by the user. By integrating advanced sensing technologies and intelligent algorithms, this panic button detection system aims to provide prompt assistance and ensure the safety of visually impaired passengers in public transportation settings.

Classification of different size of potholes based on surface area using convolutional neural network

Potholes are a significant hazard, causing severe damage to vehicles and potentially leading to fatal accidents. The automatic detection of potholes is crucial for timely maintenance and reducing these risks. This research evaluates the performance of three pre-trained Convolutional Neural Network models—ResNet 50, ResNet 18, and MobileNet—in classifying pavement images. Initially, the models distinguish between images containing potholes and those without (Potholes vs. Normal). Subsequently, they categorize pavement images into three classes: Small Pothole, Large Pothole, and Normal. Pavement images were captured from two different heights: 3.5 feet (waist height) and 2 feet. Among the models, MobileNet v2 demonstrated the highest accuracy of 98% for detecting potholes. For images taken at 2 feet, the classification accuracies for large potholes, small potholes, and normal pavement were 87.33%, 88.67%, and 92%, respectively. For images taken from waist height, the accuracies increased to 98.67%, 98.67%, and 100%, respectively. The study highlights the potential of these models in enhancing road safety through reliable and automated pothole detection, providing valuable insights for infrastructure maintenance.Article HighlightsPotholes are dangerous and can cause serious vehicle damage.Train CNN models for the classification of potholesThese CNNs, for different sizes of potholes can be used for the rehabilitation of potholes.Among all CNN modes, MobileNet v2 achieves 98% accuracy in detecting the presence of a pothole.

A Performance Evaluation of Convolutional Neural Network Architectures for Pterygium Detection in Anterior Segment Eye Images.

In this article, various convolutional neural network (CNN) architectures for the detection of pterygium in the anterior segment of the eye are explored and compared. Five CNN architectures (ResNet101, ResNext101, Se-ResNext50, ResNext50, and MobileNet V2) are evaluated with the objective of identifying one that surpasses the precision and diagnostic efficacy of the current existing solutions. The results show that the Se-ResNext50 architecture offers the best overall performance in terms of precision, recall, and accuracy, with values of 93%, 92%, and 92%, respectively, for these metrics. These results demonstrate its potential to enhance diagnostic tools in ophthalmology.

Crack analysis of tall concrete wind towers using an ad-hoc deep multiscale encoder–decoder with depth separable convolutions under severely imbalanced data

An accurate and timely cracking assessment, including the presence, location and crack geometric feature measurement, is crucial for evaluating concrete wind towers. Therefore, the early identification of cracks is a critical procedure in promptly evaluating structural integrity. This study proposed an ad-hoc encoder–decoder network based on DeepLabv3+ with depth separable convolutions to automatically segment cracks from real-world images captured from various concrete wind towers. The combined advantages of the improved DeepLabv3+ and the lightweight MobileNet v2 are suitable as a benchmark due to their high performance and universality. Four experiments were conducted to determine the model design choice and crack feature measurement capability: (1) six parametric tests using various pre-trained base networks and algorithm optimisers, (2) the influence of complex background noise (i.e., handwriting script) on crack segmentation performance, (3) comparative studies with cutting-edge pixel-wise segmentation models and (4) crack feature measurement (i.e., length and width). The research outcome demonstrated that DeepLabv3+ with MobileNet v2 can potentially be applied for efficient and accurate crack segmentation in concrete wind towers with complex backgrounds.

A novel infrared thermography image analysis for transformer condition monitoring

Electrical systems are deeply ingrained in most industrial facilities, their maintenance is increasingly becoming a critical and significant component of economic policy. Condition monitoring of electrical transformers is essential for improving their dependability and availability, averting costly maintenance and additional significant breakdowns. This research adopts the approach based on infrared thermography techniques (IRT) to keep eye on electrical transformers and detect their defects. Thermal images of the transformer were captured at two distinct operating states with an infrared camera. These images were then compiled into a dataset for further analysis. This method uses infrared imaging (IRT), along with feature analysis and machine learning, to identify issues in electrical transformers in a new way. To find the best performing machine learning model, different techniques are compared here in terms of their accuracy and stability. Two approaches are investigated for identifying features in thermography images. Approach-1 employed five common machine learning algorithms, such as Support Vector Machine (SVM), K-Nearest Neighbours (KNN), Decision Tree (DT), Logistic Regression (LR), and Least Squares Support Vector Machine (LS-SVM). Approach-2 utilized four deep learning techniques, such as MobileNetV2 (MNV2), InceptionV3 (InV3), DenseNet121(DN121), and our proposed modified VGG-16. Among all evaluated methods, the modified VGG-16 architecture achieved the highest level of dependability, demonstrating exceptional efficiency and accuracy in transformer condition monitoring and fault diagnosis.

A deep learning LSTM-based approach for AMD classification using OCT images

Age-related macular degeneration (AMD) is an age-related, persistent, painless eye disease that impairs central vision. The central area (macula) of the retina, located at the back of the eye, sustains damage that is the cause of loss of vision. The early detection of AMD can increase the probability of treatment and prevent vision loss. The AMD can be classified into dry and wet AMD based on the absence of neovascularization. This study introduces a new methodology for the classification of AMD using optical coherence tomography (OCT) retinal images. The proposed methodology is based on three stages. The first stage is the data preparation stage for resizing and normalizing the used images. The second stage is the image processing stage for enhancing the image quality as contrast and resolution these enhancements have been checked by the weighted peak signal-to-noise ratio (WPSNR) methodology. The third stage is the deep feature extraction and classification stage, which consists of two sub-models. The first model is MobileNet V1 which has been used as a deep feature extractor. The second model is LSTM (long short-term memory), fed with deep features to classify the AMD stages. A multi-classification with six separate trials has been employed with the proposed methodology, and compared with other models like DenseNet201 and InceptionV3. The proposed model has been tested on a sample of benchmark data with 4005 grayscale images labeled into three classes. The proposed methodology has achieved an accuracy of 98.85%, a sensitivity of 99.09%, and a specificity of 99.1%. To ensure the effectiveness of the proposed methodology, a comparative analysis has been established with previous approaches in the related field, and the results demonstrated the superiority of the proposed system in AMD multi-classification.

Identification of multimodal vortex optical orbital angular momentum in multimode fiber speckle patterns

Orbital angular momentum (OAM) multiplexing technology is a highly promising approach that can significantly increase the data capacity of optical communication systems. However, traditional optical communication systems are prone to atmospheric disturbances such as turbulence, which can degrade transmission quality. To mitigate this issue, multimode fibers (MMFs) have been introduced to reduce the impact of turbulence on signals. Moreover, previous studies have primarily focused on the identification of single vortex beams, facing challenges in accurately recognizing multiplexed modes. To address this challenge, this chapter proposes a multi-label image classification optimization algorithm based on transfer learning. By utilizing the pre-trained MobileNet V2 model as a feature extractor, this network structure can accurately identify 8-bit, 16-bit, and 24-bit multiplexed OAM from speckle patterns in multimode fibers, even with small sample datasets, achieving classification accuracies exceeding 95%. This method overcomes the limitations of traditional optical communication systems that are susceptible to atmospheric disturbances, providing new possibilities for long-distance transmission and increased data capacity in optical communication systems.

Adaptive Filter Pruning via Sensitivity Feedback.

Filter pruning is advocated for accelerating deep neural networks without dedicated hardware or libraries, while maintaining high prediction accuracy. Several works have cast pruning as a variant of l1 -regularized training, which entails two challenges: 1) the l1 -norm is not scaling-invariant (i.e., the regularization penalty depends on weight values) and 2) there is no rule for selecting the penalty coefficient to trade off high pruning ratio for low accuracy drop. To address these issues, we propose a lightweight pruning method termed adaptive sensitivity-based pruning (ASTER) which: 1) achieves scaling-invariance by refraining from modifying unpruned filter weights and 2) dynamically adjusts the pruning threshold concurrently with the training process. ASTER computes the sensitivity of the loss to the threshold on the fly (without retraining); this is carried efficiently by an application of L-BFGS solely on the batch normalization (BN) layers. It then proceeds to adapt the threshold so as to maintain a fine balance between pruning ratio and model capacity. We have conducted extensive experiments on a number of state-of-the-art CNN models on benchmark datasets to illustrate the merits of our approach in terms of both FLOPs reduction and accuracy. For example, on ILSVRC-2012 our method reduces more than 76% FLOPs for ResNet-50 with only 2.0% Top-1 accuracy degradation, while for the MobileNet v2 model it achieves 46.6% FLOPs Drop with a Top-1 Acc. Drop of only 2.77%. Even for a very lightweight classification model like MobileNet v3-small, ASTER saves 16.1% FLOPs with a negligible Top-1 accuracy drop of 0.03%.

Optimized Crop Disease Identification in Bangladesh: A Deep Learning and SVM Hybrid Model for Rice, Potato, and Corn.

Agriculture plays a vital role in Bangladesh's economy. It is essential to ensure the proper growth and health of crops for the development of the agricultural sector. In the context of Bangladesh, crop diseases pose a significant threat to agricultural output and, consequently, food security. This necessitates the timely and precise identification of such diseases to ensure the sustainability of food production. This study focuses on building a hybrid deep learning model for the identification of three specific diseases affecting three major crops: late blight in potatoes, brown spot in rice, and common rust in corn. The proposed model leverages EfficientNetB0's feature extraction capabilities, known for achieving rapid high learning rates, coupled with the classification proficiency of SVMs, a well-established machine learning algorithm. This unified approach streamlines data processing and feature extraction, potentially improving model generalizability across diverse crops and diseases. It also aims to address the challenges of computational efficiency and accuracy that are often encountered in precision agriculture applications. The proposed hybrid model achieved 97.29% accuracy. A comparative analysis with other models, CNN, VGG16, ResNet50, Xception, Mobilenet V2, Autoencoders, Inception v3, and EfficientNetB0 each achieving an accuracy of 86.57%, 83.29%, 68.79%, 94.07%, 90.71%, 87.90%, 94.14%, and 96.14% respectively, demonstrated the superior performance of our proposed model.

Real-Time Personal Protective Equipment Non-Compliance Recognition on AI Edge Cameras

Despite advancements in technology, safety equipment, and training within the construction industry over recent decades, the prevalence of fatal and nonfatal injuries and accidents remains a significant concern among construction workers. Hard hats and safety vests are crucial safety gear known to mitigate severe head trauma and other injuries. However, adherence to safety protocols, including the use of such gear, is often inadequate, posing potential risks to workers. Moreover, current manual safety monitoring systems are laborious and time-consuming. To address these challenges and enhance workplace safety, there is a pressing need to automate safety monitoring processes economically, with reduced processing times. This research proposes a deep learning-based pipeline for real-time identification of non-compliance with wearing hard hats and safety vests, enabling safety officers to preempt hazards and mitigate risks at construction sites. We evaluate various neural networks for edge deployment and find that the Single Shot Multibox Detector (SSD) MobileNet V2 model excels in computational efficiency, making it particularly suitable for this application-oriented task. The experiments and comparative analyses demonstrate the pipeline’s effectiveness in accurately identifying instances of non-compliance across different scenarios, underscoring its potential for improving safety outcomes.

Skin cancer classification based on an optimized convolutional neural network and multicriteria decision-making

Skin cancer is a type of cancer disease in which abnormal alterations in skin characteristics can be detected. It can be treated if it is detected early. Many artificial intelligence-based models have been developed for skin cancer detection and classification. Considering the development of numerous models according to various scenarios and selecting the optimum model was rarely considered in previous works. This study aimed to develop various models for skin cancer classification and select the optimum model. Convolutional neural networks (CNNs) in the form of AlexNet, Inception V3, MobileNet V2, and ResNet 50 were used for feature extraction. Feature reduction was carried out using two algorithms of the grey wolf optimizer (GWO) in addition to using the original features. Skin cancer images were classified into four classes based on six machine learning (ML) classifiers. As a result, 51 models were developed with different combinations of CNN algorithms, without GWO algorithms, with two GWO algorithms, and with six ML classifiers. To select the optimum model with the best results, the multicriteria decision-making approach was utilized to rank the alternatives by perimeter similarity (RAPS). Model training and testing were conducted using the International Skin Imaging Collaboration (ISIC) 2017 dataset. Based on nine evaluation metrics and according to the RAPS method, the AlexNet algorithm with a classical GWO yielded the optimum model, achieving a classification accuracy of 94.5%. This work presents the first study on benchmarking skin cancer classification with many models. Feature reduction not only reduces the time spent on training but also improves classification accuracy. The RAPS method has proven its robustness in the problem of selecting the best model for skin cancer classification.

Optimizing Convolutional Neural Networks for Image Classification on Resource-Constrained Microcontroller Units

Running machine learning algorithms for image classification locally on small, cheap, and low-power microcontroller units (MCUs) has advantages in terms of bandwidth, inference time, energy, reliability, and privacy for different applications. Therefore, TinyML focuses on deploying neural networks on MCUs with random access memory sizes between 2 KB and 512 KB and read-only memory storage capacities between 32 KB and 2 MB. Models designed for high-end devices are usually ported to MCUs using model scaling factors provided by the model architecture’s designers. However, our analysis shows that this naive approach of substantially scaling down convolutional neural networks (CNNs) for image classification using such default scaling factors results in suboptimal performance. Consequently, in this paper we present a systematic strategy for efficiently scaling down CNN model architectures to run on MCUs. Moreover, we present our CNN Analyzer, a dashboard-based tool for determining optimal CNN model architecture scaling factors for the downscaling strategy by gaining layer-wise insights into the model architecture scaling factors that drive model size, peak memory, and inference time. Using our strategy, we were able to introduce additional new model architecture scaling factors for MobileNet v1, MobileNet v2, MobileNet v3, and ShuffleNet v2 and to optimize these model architectures. Our best model variation outperforms the MobileNet v1 version provided in the MLPerf Tiny Benchmark on the Visual Wake Words image classification task, reducing the model size by 20.5% while increasing the accuracy by 4.0%.

Multiclass Classification using Enhanced MobileNet V2 Architecture

In computer vision applications including object detection, face recognition, image classification, etc., the Convolutional Neural Network (CNN) predominates. MobileNet v2 is one such CNN architecture which can significantly cut down the parameters and cost of calculation while giving comparatively higher accuracy. In this paper, we propose the enhanced MobileNet architecture which shows an improvement in the accuracy. We have applied transfer learning techniques to fine-tune the MobileNet V2 model on the food dataset. The proposed model’s generalization capacity, it’s robustness and recognition accuracy shows a considerable amount of improvement with respect to the base architecture. The findings demonstrate that the enhanced MobileNet architecture’s accuracy for food images recognition and classification has increased to 97.16%. Key Words: Convolutional Neural Network (CNN), MobileNet v2, transfer learning, image recognition and classification