Convolutional Neural Network Design Research Articles

KLASIFIKASI BUAH SEGAR DAN BUSUK MENGGUNAKAN ALGORITMA CONVOLUTIONAL NEURAL NETWORK DENGAN TFLITE SEBAGAI MEDIA PENERAPAN MODEL MACHINE LEARNING

CNN (Convolutional Neural Network) is one of the type of Deep Learning algorithm that can accept input in the form of images, or objects such as fruit, animals, office equipment, and so on. This CNN works is inspired by the neuron system in humans based on visual perception. The CNN model can be implemented into the Android system using TFLite. TFLite is a Machine Learning library specially designed for Android devices to create Machine Learning models in small binary and low latency. This can make it easier for Machine Learning researchers to use smartphones to develop the research system. Current technological advances make the fruit industry no longer use the manual fruit sorting process carried out by humans. The process of sorting fruit that is fit for consumption (fresh) and unfit for consumption (rotten) can reach hundreds and thousands, so it needs the help of machines to increase work speed. This fruit sorting process can be circumvented by creating a Machine Learning model that can classify fresh and rotten fruit which will then be processed by the system based on the desired output requirements. If the model can perform the classification well, then the fruit industry can take steps to make an automatic sorter by implementing the model into its system. This study aims to create a model with the CNN (Convolutional Neural Network) Algorithm and design a prototype system that can classify fresh and rotten fruit (red apples, bananas, and mandarin oranges) and apply the resulting model using TFLite on an Android Smartphone. The results indicated that the resulting prototype model can be used by users in the Android system. The results of the training model obtained an accuracy value of 98.20%. Meanwhile, the results of testing showed that the model can classify images from the same or different sources by 91.65% from 1090 images. These results indicate that the model is included in the excellent classification category.

Performance analysis of seven Convolutional Neural Networks (CNNs) with transfer learning for Invasive Ductal Carcinoma (IDC) grading in breast histopathological images

Computer-aided Invasive Ductal Carcinoma (IDC) grading classification systems based on deep learning have shown that deep learning may achieve reliable accuracy in IDC grade classification using histopathology images. However, there is a dearth of comprehensive performance comparisons of Convolutional Neural Network (CNN) designs on IDC in the literature. As such, we would like to conduct a comparison analysis of the performance of seven selected CNN models: EfficientNetB0, EfficientNetV2B0, EfficientNetV2B0-21k, ResNetV1-50, ResNetV2-50, MobileNetV1, and MobileNetV2 with transfer learning. To implement each pre-trained CNN architecture, we deployed the corresponded feature vector available from the TensorFlowHub, integrating it with dropout and dense layers to form a complete CNN model. Our findings indicated that the EfficientNetV2B0-21k (0.72B Floating-Point Operations and 7.1 M parameters) outperformed other CNN models in the IDC grading task. Nevertheless, we discovered that practically all selected CNN models perform well in the IDC grading task, with an average balanced accuracy of 0.936 ± 0.0189 on the cross-validation set and 0.9308 ± 0.0211on the test set.

Brain Diseases Detection and Prediction Using DeepQ Convolution Neural Network in Colab

Purpose: The paper aims to analyze the detection and prediction of brain diseases for future betterment using Convolutional neural network. Objectives: The main objective of this journal paper is to find the most correct technique of detective work in various brain diseases like Alzheimer’s disease and brain tumours using machine learning and deep learning-based approaches. Methodology: An automatic tool for neoplasm classification based on magnetic resonance imaging information is given wherever sample image slices are fed to a convolutional neural network (CNN) supported by the ResNet Squeeze and Excitation model. Alzheimer's disease misdetection system Convolutional Neural Network (CNN) design using resonance imaging (MRI) scan images. Results: Create an app-based user interface for hospitals that enables medical professionals to quickly determine the effects of tumours and Alzheimer's and recommend treatments. We can attempt and make predictions about the location and severity of mental illnesses from volume-based 3-D images because the performance and complexity of ConvNets depend on the input data visualisation. Improvements are made to surgery planning, education, and computer guidance by creating 3-D anatomical models from specific patients. Originality/Value: The results provide a brief overview of brain diseases detection and prediction with better improved form accurately. Type of Paper: Conceptual research paper.

Alpha-SGANet: A multi-attention-scale feature pyramid network combined with lightweight network based on Alpha-IoU loss.

The design of deep convolutional neural networks has resulted in significant advances and successes in the field of object detection. However, despite these achievements, the high computational and memory costs of such object detection networks on the edge or in mobile scenarios are one of the most significant barriers to their broad adoption. To solve this problem, this paper introduces an improved lightweight real-time convolutional neural network based on YOLOv5, called Alpha-SGANet: A multi-attention-scale feature pyramid network combined with a lightweight network based on Alpha-IoU loss. Firstly, we add one more prediction head to detect different-scale objects, design a lightweight and efficient feature extraction network using ShuffleNetV2 in the backbone, and reduce information loss using the SPP module with a smaller convolutional nucleus. Then, cleverly, employ GAFPN to improve feature transition processing in the neck region, including the usage of the Ghost module to construct efficient feature maps to help prediction. The CBAM module was further integrated to find areas of interest in the scene; finally, combined with Alpha-IOU loss for model supervision training, the biggest performance improvement was achieved. The experiment results show that, compared with YOLOv5s, our proposed method can achieve higher accuracy with fewer parameters and has real-time speed through verification on the PASCAL VOC dataset and MS COCO dataset.

Context-Aware, Reference-Free Local Motion Metric for CBCT Deformable Motion Compensation.

Deformable motion is one of the main challenges to image quality in interventional cone beam CT (CBCT). Autofocus methods have been successfully applied for deformable motion compensation in CBCT, using multi-region joint optimization approaches that leverage the moderately smooth spatial variation motion of the deformable motion field with a local neighborhood. However, conventional autofocus metrics enforce images featuring sharp image-appearance, but do not guarantee the preservation of anatomical structures. Our previous work (DL-VIF) showed that deep convolutional neural networks (CNNs) can reproduce metrics of structural similarity (visual information fidelity - VIF), removing the need for a matched motion-free reference, and providing quantification of motion degradation and structural integrity. Application of DL-VIF within local neighborhoods is challenged by the large variability of local image content across a CBCT volume, and requires global context information for successful evaluation of motion effects. In this work, we propose a novel deep autofocus metric, based on a context-aware, multi-resolution, deep CNN design. In addition to the inclusion of contextual information, the resulting metric generates a voxel-wise distribution of reference-free VIF values. The new metric, denoted CADL-VIF, was trained on simulated CBCT abdomen scans with deformable motion at random locations and with amplitude up to 30 mm. The CADL-VIF achieved good correlation with the ground truth VIF map across all test cases with R2 = 0.843 and slope = 0.941. When integrated into a multi-ROI deformable motion compensation method, CADL-VIF consistently reduced motion artifacts, yielding an average increase in SSIM of 0.129 in regions with severe motion and 0.113 in regions with mild motion. This work demonstrated the capability of CADL-VIF to recognize anatomical structures and penalize unrealistic images, which is a key step in developing reliable autofocus for complex deformable motion compensation in CBCT.

MobileNet and Knowledge Distillation-Based Automatic Scenario Recognition Method in Vehicle-to-Vehicle Systems

Automatic scenario recognition (ASR) based on channel status information (CSI) is an important auxiliary technology for various wireless communication systems, especially vehicle-to-vehicle (V2V) systems. CSI-based ASR methods, only based on the on-board wireless receivers rather than other expensive sensors, can identify the surrounding environments of vehicles to serve assisted and autonomous driving technologies. Recently, deep learning (DL)-based ASR methods have been proposed, which generally outperform traditional, but they generally have high computation complexity and large model sizes, which are hardly introduced into mobile edge devices, such as smart vehicles. In this paper, a lightweight convolutional neural network design method, inspired by MobileNet, is introduced into ASR for reducing computation complexity and model sizes. In addition, this method may bring about performance loss, and knowledge distillation (KD) is adopted to make up for this performance loss. Simulation results demonstrated that our proposed ASR method just has 30% computation complexity and model size of the original method, and its performance loss is very limited, especially after KD.

Multimodal convolutional neural network model with information fusion for intelligent fault diagnosis in rotating machinery

Accurate and efficient fault diagnosis in rotating machinery has long been important and challenging, as it strongly affects the system reliability and safety of industrial applications. In recent years, deep-learning-based methods are developing rapidly and have been widely used in different areas. However, most of them are data-driven and focus on the architecture and design of convolutional neural network (CNN) models, while neglecting the representation of information itself. The intrinsic characteristics of the signal can not fully explored. Moreover, rich multidirectional information hidden inside the signal, which is the key to improving the predictive performance of the entire fault diagnosis model, has usually been ignored. In this work, we propose a multimodal neural-network-based model to pursue feature representation more efficiently and effectively and further improve the diagnostic performance. This method innovatively combines continuous wavelet transform and symmetrized dot pattern graphs through the channel information fusion mechanism after the two-dimensional domain modal transformation of the time-domain signal. The integration of one- and two-dimensional convolutions could fully utilize the feature extraction capability of CNN for multimodal signals, thus forming a multimodal CNN architecture under two-level information fusion. The experiment results prove that the designed model can achieve better performance than the traditional single-modal CNN structure.

H2G-Net: A multi-resolution refinement approach for segmentation of breast cancer region in gigapixel histopathological images.

Over the past decades, histopathological cancer diagnostics has become more complex, and the increasing number of biopsies is a challenge for most pathology laboratories. Thus, development of automatic methods for evaluation of histopathological cancer sections would be of value. In this study, we used 624 whole slide images (WSIs) of breast cancer from a Norwegian cohort. We propose a cascaded convolutional neural network design, called H2G-Net, for segmentation of breast cancer region from gigapixel histopathological images. The design involves a detection stage using a patch-wise method, and a refinement stage using a convolutional autoencoder. To validate the design, we conducted an ablation study to assess the impact of selected components in the pipeline on tumor segmentation. Guiding segmentation, using hierarchical sampling and deep heatmap refinement, proved to be beneficial when segmenting the histopathological images. We found a significant improvement when using a refinement network for post-processing the generated tumor segmentation heatmaps. The overall best design achieved a Dice similarity coefficient of 0.933±0.069 on an independent test set of 90 WSIs. The design outperformed single-resolution approaches, such as cluster-guided, patch-wise high-resolution classification using MobileNetV2 (0.872±0.092) and a low-resolution U-Net (0.874±0.128). In addition, the design performed consistently on WSIs across all histological grades and segmentation on a representative × 400 WSI took ~ 58 s, using only the central processing unit. The findings demonstrate the potential of utilizing a refinement network to improve patch-wise predictions. The solution is efficient and does not require overlapping patch inference or ensembling. Furthermore, we showed that deep neural networks can be trained using a random sampling scheme that balances on multiple different labels simultaneously, without the need of storing patches on disk. Future work should involve more efficient patch generation and sampling, as well as improved clustering.

NC-Net: Efficient Neuromorphic Computing Using Aggregated Subnets on a Crossbar-Based Architecture With Nonvolatile Memory

Neuromorphic computing chips consisting of crossbar arrays of emergent nonvolatile memory (NVM) have the potential of achieving both high energy efficiency and throughput as the low-power implementation of convolutional neural network (CNN) inference engines. However, such hardware has design constraints, such as its limited fan-in/fan-out and resource-inefficient mapping, that make the design and deployment of CNN on them challenging. As a result, the user has to design the CNN model with intricate knowledge of the hardware architecture and even cannot fit the models in the hardware for CNN with high resolution image input. In this article, we propose the use of <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">aggregated subnets</i> , NC-net, which is a constrained form of the traditional layer structure, to solve these issues. With our method, we put forward an energy-efficient buffer- and analogue-to-digital converter and digital-to-analogue converter (ADC/DAC)-free architecture and a scalable end-to-end solution that automatically satisfies the hardware constraints of crossbar architectures, while optimizing the resource usage. In our solution, the exploration and deployment of a CNN for a neuromorphic crossbar hardware start with a design front end based on <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">TensorFlow</i> . Our automated design flow maps the NC-net network from <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">TensorFlow</i> to the crossbar architecture. We tested our designs on both a simulator and a field-programmable gate array (FPGA) emulator with various benchmarks. In addition to general benchmarks, including MNIST, SVHN, CIFAR-10, and CIFAR-100, we tested our system on a real-world application, human detection with high resolution (224 <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\times $ </tex-math></inline-formula> 224) images as the input. Our system achieves the state-of-the-art accuracy for these benchmarks on the crossbar-based neuromorphic hardware, with an accuracy of more than 90% for the latter. It also yielded up to <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$4.25\times $ </tex-math></inline-formula> improvement in the efficiency of spiking core usage compared to TrueNorth.

Silicon photonic secure communication using artificial neural network

Optical chaotic secure communication based on silicon photonic microcavities has the advantages of high integration capability and natural compatibility with complementary metal oxide semiconductor (CMOS) technology. To overcome the difficulties of physical device fabrication and parameter matching, we propose a hybrid scheme that combines silicon photonic microcavity (SPM) chaos and artificial neural networks (ANNs) to achieve chaotic secure communication, which could significantly enhance the feasibility of silicon photonic secure communication. We trained three typical ANNs: multilayer perceptron (MLP), long short-term memory network (LSTM), and co`nvolutional neural network (CNN). Compared with the traditional Runge-Kutta-4 function, the prediction accuracy of three networks is greatly improved. All the networks provide an R2 of 99.9 %, with root-mean-square errors (RMSEs) and mean-absolute errors (MAEs) below 10−4. Furthermore, by loading non-return-to-zero (NRZ) information and using CNN to synchronise SPM chaos, 5 Mb/s, 10 Mb/s, and 15 Mb/s NRZ encryption and decryption are all realised. This can provide synchronisation coefficients of 99.8 % and clear decrypted eye diagrams. Additionally, return-to-zero (RZ), non-return-to-zero-inverted (NRZI) information are further demonstrated and a FPGA-based CNN design is given, proving the generality of CNN decryption scheme. These results could provide a useful foundation for advancing physical and secure communication with silicon photonics chips and ANNs.

Contactless Inline IV Measurement of Solar Cells Using an Empirical Model

The current–voltage measurement is the most important measurement in solar cell quality control. As the contacting process of cells results in mechanical stress and consumes a significant amount of measurement time, this work presents an IV characterization based on contactless measurements only. An empirical model is introduced that can derive the full IV curve and IV parameters as the open‐circuit voltage, short‐circuit current density, fill factor, and efficiency. As a basis, a series of photoluminescence and contactless electroluminescence images and spectral reflectance measurements are used. An advantage of the model's convolutional neural network design lies in the semantic compression of local image structures across the input data. Within an ablation study, it is shown that the empirical model is well suited to combine these data sources, which is the optimal input configuration for contactless IV derivation. The accuracy, e.g., with an error in efficiency of and correlation of over 99%, is similar to comparing two contacting IV measurement devices. The contactless IV curves also have a close fit to their contacted counterparts. Within simulations on module level, it is demonstrated that contactless binning performs as well as contacting binning and does not result in any additional mismatch loss.

Designing rotationally invariant neural networks from PDEs and variational methods

Partial differential equation models and their associated variational energy formulations are often rotationally invariant by design. This ensures that a rotation of the input results in a corresponding rotation of the output, which is desirable in applications such as image analysis. Convolutional neural networks (CNNs) do not share this property, and existing remedies are often complex. The goal of our paper is to investigate how diffusion and variational models achieve rotation invariance and transfer these ideas to neural networks. As a core novelty, we propose activation functions which couple network channels by combining information from several oriented filters. This guarantees rotation invariance within the basic building blocks of the networks while still allowing for directional filtering. The resulting neural architectures are inherently rotationally invariant. With only a few small filters, they can achieve the same invariance as existing techniques which require a fine-grained sampling of orientations. Our findings help to translate diffusion and variational models into mathematically well-founded network architectures and provide novel concepts for model-based CNN design.

Design of a quantum convolutional neural network on quantum circuits

This paper introduces a quantum convolutional neural network model that is implementable on real quantum circuits. Three basic blocks, including the quantum encoding block, the model design block and the parameter tuning block, are designed to formulate the quantum convolutional neural network framework. Simulation results on the MNIST dataset are presented to demonstrate the representation, learning, and optimization capacity of the proposed model.

PCUNet: A Context-Aware Deep Network for Coarse-to-Fine Point Cloud Completion

Point cloud completion aims at predicting a complete 3D shape from an incomplete input. It has important applications in the fields of intelligent manufacturing, augmented reality, virtual reality, self-driving cars, and intelligent robotics. Although deep learning-based point cloud completion technology has developed rapidly in recent years, there are still unsolved problems. Previous approaches predict each point independently and ignore contextual information. And, they usually predict a complete 3D shape based on a global feature vector extracted from an incomplete input, which leads to missing of some fine-grained details. In this paper, motivated by the transposed convolution and the “UNet” structure in neural networks for image processing, we propose a context-aware deep network termed as PCUNet for coarse-to-fine point cloud completion. It adopts an encoder-decoder structure, in which the encoder follows the design of the relation-shape convolutional neural network (RS-CNN), and the decoder consists of fully-connected layers and two stacked decoder modules for predicting complete point clouds. The contributions are twofold. First, we design the decoder module as a coordinate-guided context-aware upsampling module, in which contextual information can be taken into full account by neighbor aggregation. Second, to preserve fine-grained details in the input, we propose attention-enhanced skip connections for effective information propagation from the encoder to the decoder. Experiments are conducted on the widely used PCN and KITTI datasets. The results show that our proposed approach achieves competitive performance compared to the existing state-of-the-art approaches in terms of the Chamfer distance and the computational complexity metrics.

Assessing the effects of convolutional neural network architectural factors on model performance for remote sensing image classification: An in-depth investigation

Although the application of deep learning in remote sensing (RS) has achieved fruitful results, systematic research on exploring the model performance and guiding the design of new convolutional neural network (CNN) architectures is still lacking. This subject is of great concern to researchers or practitioners in this field because existing CNN structures may not be adequate to deal with complex RS scenarios. In this study, an empirical formula of CNN model performance is delivered based on a literature review. Extensive experiments are conducted on six public RS data sets to investigate the influences of three architectural factors, namely, network depth, width, and cardinality. Two types of CNN architectures, i.e., VGG and ResNet, are adopted as baselines. We monitor and visualize the data distributions and gradients of the utilized CNNs to prevent the gradient vanishing or exploding problem. Grad-CAM is adopted to open the black box of CNNs and to illustrate the effects of adjusting architectural factors. Our experiments indicate that (1) increasing the network depth is beneficial to the semantic feature learning capacity of a CNN model, but excessive depth also leads to a decline of overall accuracy; (2) a partly widening strategy is effective because it can improve the model performance while maintaining the network complexity; and (3) network cardinality has huge potential in achieving a balance between model efficiency and accuracy. Suggestions for improving the CNN model performance and developing new structures are summarized in this paper.

Computational Analysis of stochastic arithmetic computing and stochastic activation function

Stochastic Number Generators (SNG) plays a significant role in designing a stochastic computing system. SNGs make the stochastic system comfortable for computing in the stochastic domain. The challenges in developing the stochastic computing system are correlation and hardware area occupancy. By considering these phenomena, we have considered Linear Feedback Shift Register (LFSR) based SNG and S-box based SNG in this work. Our contributions to this paper are stochastic computation for activation functions using the SNGs mentioned above and stochastic computation for arithmetic components in the stochastic domain. By considering the two SNG methods, the difference in the computation for accuracy has been analyzed for stochastic activation function and stochastic arithmetic computation. The better SNG method will be used as SNG for stochastic convolutional neural network design using this analysis.

Transfer Learning-Based Multi-Scale Denoising Convolutional Neural Network for Prostate Cancer Detection.

Simple SummaryTo enhance the automatic diagnosis of the prostate cancer using machine learning algorithm, we modify the design of convolutional neural network to support multi-scale denoising of cancer images. Transfer learning is employed to leverage the detection accuracy of the prostate cancer detection model by taking advantages from more unseen data from a source dataset. Compared to existing methodologies, our work improves the accuracy by more than 10%. Ablation studies have conducted to evaluate the contributions of the components of the proposed algorithm, with 2.80%, 3.30%, and 3.13% for image denoising, multi-scale scheme, and transfer learning, respectively. The results reveal the effectiveness of the algorithm and provide insights for five future research directions.Background: Prostate cancer is the 4th most common type of cancer. To reduce the workload of medical personnel in the medical diagnosis of prostate cancer and increase the diagnostic accuracy in noisy images, a deep learning model is desired for prostate cancer detection. Methods: A multi-scale denoising convolutional neural network (MSDCNN) model was designed for prostate cancer detection (PCD) that is capable of noise suppression in images. The model was further optimized by transfer learning, which contributes domain knowledge from the same domain (prostate cancer data) but heterogeneous datasets. Particularly, Gaussian noise was introduced in the source datasets before knowledge transfer to the target dataset. Results: Four benchmark datasets were chosen as representative prostate cancer datasets. Ablation study and performance comparison between the proposed work and existing works were performed. Our model improved the accuracy by more than 10% compared with the existing works. Ablation studies also showed average improvements in accuracy using denoising, multi-scale scheme, and transfer learning, by 2.80%, 3.30%, and 3.13%, respectively. Conclusions: The performance evaluation and comparison of the proposed model confirm the importance and benefits of image noise suppression and transfer of knowledge from heterogeneous datasets of the same domain.

A convolutional neural network based crystal plasticity finite element framework to predict localised deformation in metals

Convolutional neural networks (CNNs) find vast applications in the field of image processing. This study utilises the CNNs in conjunction with the crystal plasticity finite element method (CPFEM). This research presents a framework that enables CNNs to make rapid and high-fidelity predictions for materials under uniaxial tension loading. The inputs to the CNN model are material hardening parameters (initial hardness and initial hardening modulus), a global tensile strain value, and microstructure with a varying number of grains, grain size, grain morphology and texture. This input selection allows performing simulations for a wide range of materials, as defined by microstructure and flow curves. The outputs of the CNN are the local stress and strain values. The proposed framework involves the following stages: feature engineering, generation of synthetic microstructures, CPFEM simulations, data extraction and preprocessing, CNN design and selection, CNN training, and validation of the trained network. The trained CNN was successfully demonstrated to predict local stress and strain evolution for the completely new dataset (test set) containing synthesised microstructures. The test set predictions were evaluated, and the median, worst, and best predictions were presented and discussed. Overall, the CNN demonstrated excellent agreement with CPFEM simulations, thus validating its accuracy. Then, the CNN was applied to predict the stress and strain evolution for AA5754 and AA6061 microstructures obtained using electron backscatter diffraction. These two microstructures were entirely new for the CNN and displayed size and grain morphology different from the synthesised microstructures. For both microstructures, the obtained stress and strain evolution predictions demonstrated excellent agreement with CPFEM simulations, thus confirming the flexibility of the trained CNN model. Then, the framework was extended to predict strain localisation and was evaluated on an AA6061 microstructure. The results presented in this research demonstrate a clear computational advantage of CNN without loss of accuracy. Finally, the research offers prospects for future advances.

Design and development of novel hybrid optimization-based convolutional neural network for software bug localization

Design and development of novel hybrid optimization-based convolutional neural network for software bug localization

Evaluation of Diverse Convolutional Neural Networks and Training Strategies for Wheat Leaf Disease Identification with Field-Acquired Photographs

Tools for robust identification of crop diseases are crucial for timely intervention by farmers to minimize yield losses. Visual diagnosis of crop diseases is time-consuming and laborious, and has become increasingly unsuitable for the needs of modern agricultural production. Recently, deep convolutional neural networks (CNNs) have been used for crop disease diagnosis due to their rapidly improving accuracy in labeling images. However, previous CNN studies have mostly used images of single leaves photographed under controlled conditions, which limits operational field use. In addition, the wide variety of available CNNs and training options raises important questions regarding optimal methods of implementation of CNNs for disease diagnosis. Here, we present an assessment of seven typical CNNs (VGG-16, Inception-v3, ResNet-50, DenseNet-121, EfficentNet-B6, ShuffleNet-v2 and MobileNetV3) based on different training strategies for the identification of wheat main leaf diseases (powdery mildew, leaf rust and stripe rust) using field images. We developed a Field-based Wheat Diseases Images (FWDI) dataset of field-acquired images to supplement the public PlantVillage dataset of individual leaves imaged under controlled conditions. We found that a transfer-learning method employing retuning of all parameters produced the highest accuracy for all CNNs. Based on this training strategy, Inception-v3 achieved the highest identification accuracy of 92.5% on the test dataset. While lightweight CNN models (e.g., ShuffleNet-v2 and MobileNetV3) had shorter processing times (&lt;0.007 s per image) and smaller memory requirements for the model parameters (&lt;20 MB), their accuracy was relatively low (~87%). In addition to the role of CNN architecture in controlling overall accuracy, environmental effects (e.g., residual water stains on healthy leaves) were found to cause misclassifications in the field images. Moreover, the small size of some target symptoms and the similarity of symptoms between some different diseases further reduced the accuracy. Overall, the study provides insight into the collective effects of model architecture, training strategies and input datasets on the performance of CNNs, providing guidance for robust CNN design for timely and accurate crop disease diagnosis in a real-world environment.