Comparative evaluation of deep learning architectures for brinjal fruit disease classification

Comparative evaluation of deep learning architectures for brinjal fruit disease classification Basab Nath1, Yonis Gulzar2, Sagar Tamang3, Mohannad Alkanan2 1 School of Computer Science Engineering and Technology, Bennett University, Greater Noida, India 2 Department of Management Information Systems, College of Business Administration, King Faisal University, Al-Ahsa, 31982, Saudi Arabia 3 Centre for Educational Technology, Indian Institute of Technology Patna, Bihta, 835217, India Corresponding authors: Basab Nath (basab.na[email protected]); Yonis Gulzar ([email protected]) Academic editor: Yongming Liu♦Received 24 September 2025♦Accepted 6 November 2025♦Published 9 December 2025 Abstract Micronutrient malnutrition, especially anaemia, remains a pressing global concern, underscoring the value of affordable and nutrient-dense crops such as brinjal (Solanum melongena L.). Yet, its cultivation is hampered by fruit diseases that significantly diminish both yield and market quality. Conventional diagnosis depends on manual inspection, which is time-consuming and prone to error. To overcome this limitation, we conducted a comprehensive benchmarking of contemporary deep learning architectures for automated brinjal fruit disease recognition under real-world field conditions. For this purpose, we developed the BrinjalFruitX dataset containing 3,077 images of five classes—Healthy, Phomopsis Blight, Wet Rot, Shoot and Fruit Borer, and Fruit Cracking—captured under natural variability. We evaluated ten representative convolutional neural networks (CNNs), including models from the VGG, ResNet, Inception, EfficientNet, and MobileNet families, across four training paradigms: training from scratch, transfer learning, fine-tuning, and full training (“Full Monty”). Performance was systematically analyzed using accuracy, macroand weighted-F1 scores, training loss, confusion matrices, and computational efficiency indicators such as parameter count, FLOPs, and inference latency. Among the tested models, MobileNetV2 with Full Monty training achieved the best balance of performance and efficiency, reaching 97.98% accuracy, a macro-F1 of 0.9793, and operating with only 3.4M parameters, 0.30B FLOPs, and an inference time of 3.2 ms per image. While InceptionV3 and VGG16 also produced competitive results, they required considerably higher computational resources. In contrast, deeper ResNets and EfficientNetB0 offered inferior accuracy despite higher complexity. These findings highlight MobileNetV2 with full training as a practical and lightweight solution for on-device and farmer-oriented applications. This study establishes a strong benchmark for advancing deep learning–driven disease detection in sustainable agricultural systems. Keywords Brinjal fruit disease classification, convolutional neural networks (CNNs), transfer learning, model efficiency, MobileNetV2 Introduction Micronutrient malnutrition, also known as hidden hunger is a worldwide health problem. Thirty point seven percent of women of reproductive age have anaemia, while the prevalence is even held at 35.5% among pregnant women based on most recent data released by World Health Organisation (WHO, 2025). Prevalence of anaemia among children aged 6–59 months was 39.8% in 2019. Anaemia causes fatigue, decreased productivity, and bad maternal-child health outcomes — passed on across generations as poor motor and cognitive Copyright Basab Nath, et al. This is an open access article distributed under the terms of the Creative Commons Attribution License (CC-BY 4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Emirates Journal of Food and Agriculture 37: 1–14 doi: 10.3897/ejfa.2025.172982 RESEARCH PAPER SPECIAL COLLECTION NAME Basab Nath, et al.: Comparative study of CNNs for brinjal disease detection2 Emirates Journal of Food and Agriculture development (World Health Organization 2025). The world now is not on course to reach the target of cutting anemia by half by 2030. Vegetables are a critical component of human diets, as they possess necessary vitamins, minerals, phytonutrients and dietary fiber. Yet, in most areas, vegetable intake is still less than the recommended amount (Afshin et al. 2019). Crops, like brinjal (eggplant, Solanum melongena L.) and its wild relatives are important in providing dietary diversity for combating hidden hunger. Eggplant fruits are a cheap source of calcium, phosphorus, iron and B-group vitamins. They also include polyunsaturated fatty acids that reduce cholesterol, thereby adding nutritional value (Food and Agriculture Organization of the United Nations 2022). Brinjal is one of the globally important solanaceous vegetable crops. It covers 18.94 lakh hectares during 2022–23 with the production of 593 lakh tonnes an average yield of 31,383 kg/ha. This makes the brinjal as the fifth most valuable Solanaceae crop after potato, tomato, pepper and tobacco (Food and Agriculture Organization of the United Nations 2022). China is the dominant producer with 38.31 million tonnes (64.5% of worldwide production), followed by India with 12.76 million tonnes (21.5%). Egypt, Turkey and Indonesia are also major producers. Brinjal is one of the first five economically and socially most important vegetable crops in Asia and the Mediterranean. In India, the annual crop production is approximately 12.7 million tonnes signifying its significant role in the rural livelihood and national supply of vegetables (Professor Jayashankar Telangana State Agricultural University 2025). Growing is mainly in places like West Bengal, Odisha, Gujarath and Madhya Pradesh. Brinjal: The total area under brinjal in 2023–24 according to the third advance estimates was 6.78 lakh hectares with a production of 128.79 lakh tonnes, a little less than last year (Food And Agriculture Organization Of The United Nations 2022). For instance in Telangana, the price forecast for Yasangi (Rabi) 2024–25 pre-harvest estimated the range of Rs. 1720–2070 per quintal, highlighting both the economic importance and vulnerability of crops to relative production and market conditions (Professor Jayashankar Telangana State Agricultural University 2025). As for Africa, African eggplant (Solanum aethiopicum) is one of the traditional vegetables in East and West Africa. It is grown for fruits as well as leaves and has high phenotypic variation in size, colour and taste. The crop does best under hot and dry conditions and features in the local people’s diet as well as farming system (Mvungi et al. 2025). Nevertheless, it has been overlooked in socioeconomic studies and commercial crop breeding schemes except being used as a resistance donor to S. melongena. Breeding efforts by the World Vegetable Center and national institutions have resulted in the development of new improved cultivars with higher yields and decreased bitter taste levels; adopting these is geographically limited (Mvungi et al. 2025). Despite its economic and nutritional significance, brinjal cultivation faces serious challenges from fruit diseases. Common problems include shoot and fruit borer, wet rot, fruit cracking, and phomopsis blight. These diseases reduce yield, affect fruit quality, and lead to economic losses. Traditional monitoring depends on manual inspection, which is time-consuming and often unreliable. In recent years, artificial intelligence (AI) and deep learning (DL) have been introduced as effective solutions for plant disease detection (Gulzar 2025a, 2025b). However, most prior studies focus on controlled datasets of leaves, such as PlantVillage, while research on fruit disease detection remains limited. Targeted studies on brinjal fruits are scarce, and systematic evaluations of advanced architectures such as convolutional neural networks (CNNs) and vision transformers (ViTs) are yet to be reported (Professor Jayashankar Telangana State Agricultural University 2025). This study seeks to address this gap through a comprehensive benchmark of state-of-the-art deep learning models for brinjal fruit disease classification. Special emphasis is placed on class imbalance and efficiency to support robust, real-world applications in agriculture. The major contributions of this work are as follows: 1. Development of the BrinjalFruitX dataset, a curated collection of fruit images captured under realistic field conditions with natural variability. 2. Systematic benchmarking of ten widely used CNN architectures (VGG, ResNet, Inception, EfficientNet, and MobileNet) under four distinct training strategies. 3. Comprehensive evaluation using balanced classification metrics (accuracy, macro-/weighted-F1, confusion matrices) alongside efficiency indicators (parameters, FLOPs, inference time). 4. Recommendations for deploying efficient yet accurate models for farmer-level and mobile applications. Related work Plant disease detection with deep learning Deep learning has transformed plant pathology research by enabling automatic disease detection from images. Early works demonstrated the capability of convolutional neural networks (CNNs) to identify leaf diseases in crops such as sunflower (Gulzar et al. 2023), soybean (Gulzar 2024), corn (Alkanan and Gulzar 2024), and rice (Seelwal et al. 2024) with accuracies exceeding 90%. Large controlled datasets, such as PlantVillage, have accelerated progress, but their laboratory conditions limit robustness in real fields. For brinjal, where fruit symptoms occur under variable lighting and occlusion, these limitations are even more critical. Thus, while CNNs remain the dominant architecture in plant disease detection, their generalization to fruit diseases in real-world settings remains underexplored. Emir. J. Food Agric ⋅ Volume 37 ⋅ 2025 3 Emirates Journal of Food and Agriculture Eggplant and brinjal disease studies Most prior work on eggplant has concentrated on leaf diseases. Krishnaswamy and Purushothaman (Krishnaswamy et al. 2020) created a five-class dataset and used VGG16 with multi-class SVM, achieving 99.4% accuracy. Abisha et al. (Abisha et al. 2023) combined Shearlet transforms with CNNs to capture texture and structural cues, reaching 93% mean accuracy. Chelladurai and Sujatha (Chelladurai and Sujatha 2024) proposed DenseNet ensembles on a seven-class dataset of 8,080 leaf images, achieving 94.4%. These studies confirm the promise of CNNs for eggplant leaves, but they rely on controlled leaf images and do not address fruit conditions. Some studies considered both leaves and fruits. Haque and Sohel (Haque and Sohel 2022) fused CNN-SVM and CNN-Softmax pipelines for nine eggplant diseases, including fruit rot. While accuracy improved over standard models, most of the dataset still comprised leaf images. Venkataramana et al. (Attada et al. 2022) mixed CNNs with SVMs for brinjal disease prediction, reporting 99% accuracy, though validation details were limited. Collectively, these works illustrate the leaf-centric bias in the literature, with fruit disease detection still rare. Detection and robotics approaches Object detection methods have been adopted for field use. Nasution and Kartika (Nasution and Kartika 2022) applied YOLOv4 on Raspberry Pi with Telegram notifications, demonstrating feasibility for low-cost, real-time detection. Huang et al. (Huang et al. 2024) introduced YOLOv8-E, enhancing small-object detection and model compactness. Tamilarasi et al. (Tamilarasi et al. 2025) developed YOLOv11s-Brinjal, optimized for harvesting robots, reporting 98% mAP with only 8.2 MB model size. These works highlight YOLO’s suitability for real-time detection, yet they target fruit presence or pests, not disease classification on fruits. Robotics and pest monitoring also appear in parallel. Sutayco et al. (Sutayco et al. 2025) developed a solar-powered rover for insect detection that reached 86% accuracy across classes. While valuable for integrated pest management, these pipelines focus on insect categories rather than disease phenotypes. Multimodal and advanced sensing Beyond RGB imaging, multispectral and hyperspectral sensing have been tested. Zhang et al. (Zhang et al. 2024) fused five channels to detect early Verticillium wilt with ~ 87% precision. Wang et al. (Wang et al. 2024) used hyperspectral imaging and autoencoders for leaf nitrogen estimation, achieving R2 ≈ 0.91. Huang et al. (Huang et al. 2025) combined VIS-NIRS spectra with CNN+attention+LSTM to predict fruit maturity (R2 = 0.876). These studies prove the potential of spectral signals for plant monitoring but are specialized, expensive, and less accessible for farmer-level solutions. Trends in architectures The reviewed studies confirm the dominance of convolutional neural networks and YOLO-based detectors for eggplant disease and pest identification. CNNs, including VGG, ResNet, and DenseNet, have been applied extensively for classification tasks, often achieving high accuracies under controlled conditions (Krishnaswamy et al. 2020; Abisha et al. 2023; Chelladurai and Sujatha 2024). YOLO adaptations have proven effective for real-time detection in field settings, with lightweight variants demonstrating strong potential for mobile and robotic deployment (Nasution and Kartika 2022; Huang et al. 2024; Tamilarasi et al. 2025). Hybrid models that integrate fusion strategies or attention modules have also emerged, but attention has mostly been embedded as a component within CNN frameworks rather than explored through transformer architectures (Haque and Sohel 2022; Wang et al. 2025). Consequently, systematic comparisons between CNNs and vision transformers are still absent in the context of eggplant disease recognition. Taken together, prior work illustrates notable progress in leaf-based disease detection and object detection pipelines, but fruit disease classification has received little focused attention. While CNNs and YOLO variants dominate the field, issues such as class imbalance, robustness under field variability, efficiency on edge devices, and model interpretability remain insufficiently addressed. These gaps provide the rationale for a comprehensive benchmark of state-of-the-art CNN and transformer architectures for brinjal fruit disease detection under realistic conditions. Materials and methods This section describes the methodological framework used in this study which consists of an overview of dataset, preprocessing techniques, experimental setups and the deep learning models applied for brinjal fruit classification. The objective was to enable fair and reproducible comparison between several modern CNN architectures such that the architecture bias can be minimized. To that end, we developed an experimental pipeline which includes baseline benchmarking and hyperparameter optimization as well as a study on advanced training strategies. Flowchart, demonstrating the overall approaches adopted in the present work is shown in Fig. 1 which is organized into three central stages: 1. Data Acquisition & Pre-processing, which includes dataset preparation and balancing; 2. Proposal of an Experimental Framework, showing the multi-phase model assessment (Baseline Benchmarking, Hyperparameter Tuning and Overall Strategy Assessment); and 3. Results & Analysis, where the performance results are discussed and the best model is selected. Basab Nath, et al.: Comparative study of CNNs for brinjal disease detection4 Emirates Journal of Food and Agriculture Dataset and preprocessing In this paper, we employed the BrinjalFruitX (Hasan et al. Bijoy 2025) dataset consisting of brinjal fruit images taken in their original field environments. The data set is representative of real agricultural environment since the images are with different scale, illumination and orientation as well as varying background complexity. This diversity enabled us to build models that are not just accurate but also resilient in the face of real-world challenges. Each photo in the dataset was labelled to various class referring to different state of brinjal fruits. Sample images from all the classes are shown in Fig. 2. In order to make the dataset ready for training, we introduced a standardized preprocessing pipeline. We began by loading all images with OpenCV image library and then converting the loaded BGR images into an RGB format that are compatible with deep learning tools like TensorFlow and Keras. Each image was then resized to meet the input size of structures as follows: 224224 pixels for MobileNet, VGG, ResNet and EfficientNetB0; 299299 pixels for InceptionV3 and InceptionResNetV2; 600600 pixels (large-scale experiments). This rescaling provided all models with a constant range and allowed the results to be compared. Then we rescaled the pixel values of images from [0, 255] to [0, 1]. This helped to reduce the instability of the gradient updates and increased convergence speed during training. In this study the dataset was divided into three parts (70% training, 10% validation and 20% test) for experimental consistency. We preserved the class distribution in each split by stratified sampling. We did not use this test set for training or validation at any stage, and used it only for a final unbiased evaluation of model performance. As the dataset showed a modest class imbalance, we took two complementary approaches to deal with it. One way was to keep the original class distribution but to add class weights for loss calculation (Table 2). In the second method, we built a balanced sample by oversampling the minority classes through augmentation such that all categories had around same number of data_points (Table 1). Even though the current study uses publicly available BrinjalFruitX data set where images were already captured in various field conditions. We used data augmentation to Figure 1. Workflow for brinjal fruit disease classification. Figure 2. Sample images from the BrinjalFruitX dataset. Emir. J. Food Agric ⋅ Volume 37 ⋅ 2025 5 Emirates Journal of Food and Agriculture increase real-world generalisation. For the augmentation pipeline we used random horizontal flips, small angle rotations (up to ±20°), zoom in-out variations (±10%) and also brightness-contrast adjustments (±15%), all applied stochastically per mini-batch. Small translation shifts (≤10%) and shear transforms were also applied to mimic random fluctuations in natural fruit orientation and framing across the existing data set. These operations virtually expanded the diversity of training samples and reinforced model’s resistance to illumination, orientation and scale changes. The final class-wise distribution of images across the training, validation, and testing splits is summarized in Table 1. Every row represents the particular brinjal disease class. The columns under Training (Original), Validation, and Testing indicate the dataset distribution before oversampling. The Training (Balanced) column represents the number of samples gained after oversampling the minority class by data augmentation. The Total (Original) column presents the total number of real images per class before oversampling and in Total (New) the total after oversampling. Accordingly, the total number of original and augmented images in the data set was augmented from 1,802 to 3,077. Training protocols Stratified sampling was performed while splitting the BrinjalFruitX dataset to obtain 70% training, 10% validation and 20% testing sets for keeping consistency and reproducibility. This helped to maintain the relative distribution of classes and reduced bias toward majority groups. The validation set was used only for hyperparameter tuning/early stopping procedures and the test set remained completely hidden until final evaluation. Pilot 5-fold cross-validation was also performed to check for robustness but results shown in this study are the ones from fixed split, in order to facilitate comparison. We conduct all experiments on NVIDIA RTX A6000 GPUs. Random seeds were fixed for Python, NumPy and TensorFlow to improve reproducibility. The pre-processing pipeline was common to all models so that the models being compared were trained in identical conditions. Apart from this default protocol, four additional training procedures were implemented to investigate the impact of initialization, pre-training, fine-tuning and dataset balancing. These are presented in Table 2 and an overview of the entire protocol is given in Table 3. Models selected In this work, we considered a total of ten convolutional neural network (CNN) architectures from five different families: VGG, ResNet, Inception, EfficientNet and MobileNet. This variety allowed us to perform a thorough comparison of the classical deep networks with recent light-weight architectures focused on efficiency. Representative structures from each of the families are shown in Figs 3–7, and their input patterns are listed in Table 4. The VGG family (e.g., VGG16 and VGG19) (Simonyan and Zisserman 2015) was one of the earliest successes in Table 1. Class-wise dataset distribution of images before and after oversampling. Class Training (Original) Training (Balanced) Validation Testing Total images (Original) Total images (New) Brinjal Fruit Creaking 140 507 20 40 200 567 Healthy Brinjal 359 507 52 103 514 662 Phomopsis Blight 113 507 16 32 161 555 Shoot and Fruit Borer 507 507 73 145 725 725 Wet Rot 141 507 20 41 202 568 Total 1260 2535 181 361 1802 3077 Table 2. Training strategies adopted for model evaluation. Strategy Name Initial Weights Base Model State Key Techniques Applied From Scratch Random (None) Fully Trainable End-to-end training from random initialization on the balanced dataset. Transfer Learning ImageNet Fully Frozen Single-stage training of only the classification head. Fine-Tune ImageNet Stage 1: Frozen Stage 2: Top 50% Trainable Two-stage training with class weights and a low learning rate during fine-tuning. Full Monty ImageNet Stage 1: Frozen Stage 2: Top 50% Trainable Similar two-stage training on the balanced dataset, eliminating the need for class weights. Table 3. Summary of the training protocol used across all experiments. Parameter Configuration Dataset Split 70% training, 10% validation, 20% testing (stratified) Validation Strategy Fixed split (pilot 5-fold CV for robustness) Optimizer Adam Learning Rates Tested 1 × 10–3, 1 × 10–4, 1 × 10–5 Batch Sizes Tested 8, 16, 32 Max Epochs 100 Callbacks EarlyStopping, ModelCheckpoint, ReduceLROnPlateau Hardware NVIDIA RTX A6000 GPU Reproducibility Fixed random seeds across all libraries Basab Nath, et al.: Comparative study of CNNs for brinjal disease detection6 Emirates Journal of Food and Agriculture deep CNN design. These networks consist of sequential stacks of 33 convolutional layers which are followed by ReLU activations and max-pooling layers, and then fully connected for classification. Despite being very computationally intensive in terms of the number of parameters, they are still excellent baselines for extracting features. Fig. 3 is used in the implementation. The family of ResNets (ResNet50 and ResNet152) (He et al. 2016) presented residual learning by shortcut-connections, which allows very deep models without vanishing gradients. ResNet50 is equipped with 50 layers and relies on bottleneck residual blocks; whereas ResNet152 addresses the representational power issue by scaling the architecture up to 152 layers at the expense of heavier computation. Fig. 4 shows the ResNet50 architecture used in this paper. The family of inception, such as InceptionV3 and InceptionResNetV2, concentrate on efficient multi-scale feature representation. InceptionV3 (Szegedy et al. 2016) employs factorized convolutions, auxilary classifiers, and reduction modules, whereas InceptionResNetV2 (Szegedy et al. 2017) proposes to stack inception modules with residual connections for further and better model. The trimmed architecture of InceptionV3 in our experiment is shown in Fig. 5. The EfficientNet (including the smallest model, EfficientNetB0) are a family of models constructed in a similar fashion as follows: balancing network depth, width and resolution (Tan and Le 2019). Leveraging MBConv blocks with squeeze-and-excitation (SE) mechanism, EfficientNetB0 strikes a good balance between accuracy and efficiency. The overview of EfficientNetB0 is presented in Fig. 6. Last, the MobileNet family (MobileNetV2 and MobileNetV3Large) (Sandler et al. 2018; Howard et al. 2019) is designed to be lightweight. MobileNetV2 utilizes depthwise separable convolutions and inverted residual blocks with linear bottlenecks to reduce computation cost a lot. MobileNetV3Large enhances it with NAS and SE module. The architecture of the MobileNetV2 is represented in Fig. 7. Table 4 summarizes the models and their standard input sizes used in this study. Table 4. Summary of CNN models selected for evaluation, showing their optimal input size used in the final experiments. Model Architecture Family Optimal Input Size (pixels) VGG16 VGG 224 × 224 VGG19 VGG 224 × 224 ResNet50 ResNet 224 × 224 ResNet152 ResNet 224 × 224 InceptionV3 Inception 299 × 299 InceptionResNetV2 Inception 299 × 299 EfficientNetB0 EfficientNet 224 × 224 MobileNetV2 MobileNet 224 × 224 MobileNetV3Large MobileNet 224 × 224 Figure 3. VGG16 architecture with stacked 3 × 3 convolutional layers. Figure 4. Compact schematic of the ResNet50 architecture. Figure 5. Schematic of the InceptionV3 architecture with inception and reduction modules. Figure 6. Schematic of the EfficientNetB0 architecture with MBConv blocks. Figure 7. Schematic of the MobileNetV2 architecture with depth wise separable convolutions and inverted residuals. Emir. J. Food Agric ⋅ Volume 37 ⋅ 2025 7 Emirates Journal of Food and Agriculture Experimental environment settings and performance evaluation metrics All models were implemented in TensorFlow 2.x with the Keras API. We adopted the Adam optimizer, which adaptively tunes learning rates and is widely effective across CNN architectures. Three candidate learning rates (1 × 10–3, 1 × 10–4, 1 × 10–5) were tested; 1 × 10–4 consistently offered the best balance of convergence speed and stability. Batch sizes of 8, 16, and 32 were explored, with smaller values (8 or 16) generally producing better generalization. All models were trained for 100 epochs with common hyper-parameters to allow for direct comparison between architectures. The identical optimization, learning rate schedule, and augmentation pipeline were used consistently across models. We saved model checkpoints after each epoch and then used the best weights according to validation loss for final testing. The performance of all models was assessed using a comprehensive set of evaluation metrics that provide both overall accuracy and class-level insights. Since the BrinjalFruitX dataset exhibited moderate class imbalance, we emphasized not only accuracy but also precision, recall, and F1-score, along with their macroand weighted-averaged forms (Sokolova and Lapalme 2009; Powers 2011). In addition, the test loss was reported using sparse categorical cross-entropy (Goodfellow et al. 2016) to capture overall generalization performance. Classification metrics: Accuracy was computed as the ratio of correctly classified samples to the total number of samples: where TR, TN, FP, and FN represent true positives, true negatives, false positives, and false negatives, respectively. Precision and recall were calculated for each class as: , The F1-score, which balances precision and recall, was defined as: To account for multiple classes, we computed two types of F1 aggregation. The macro-averaged F1-score was obtained by averaging the F1-scores across all classes: where N is the number of classes. This treats all classes equally, regardless of their frequency. The weighted-averaged F1-score was also calculated, in which class-specific F1-scores were weighted by their support (number of true instances per class), thus reflecting the influence of class imbalance. Loss metric: The primary loss function was sparse categorical cross-entropy, given by: where yi is the ground-truth class indicator and is the predicted probability for class i (Goodfellow et al. 2016). Confusion matrix: In addition to scalar metrics, we also analyzed the confusion matrix (Kohavi and Provost 1998) for each trained model to visualize per-class performance. This provided insights into systematic misclassifications and helped identify visually similar categories where the models struggled. Efficiency metrics: To assess the practical viability of each model, we also recorded efficiency-oriented measures including the number of trainable parameters, floating point operations (FLOPs) per forward pass, and average inference time per image (batch size = 1) (Howard et al. 2017; Molchanov et al. 2017). These metrics are crucial for determining the suitability of models for deployment in resource-constrained agricultural settings. Results and discussion In this section, we evaluated our approach using the several phase framework described in material and methods section. The main findings are organized in 3 stages of validating results: (1) baseline benchmarking for setting an initial context, (2) hyperparameter tuning for adjusting training dynamics and (3) thorough model/strategy evaluation. In each case, we point out some observations that helped during search and eventually converge to the most effective architecture and training strategy. We have trained all the models with 100 epochs to maintain the balance conditions for all the experiments. Phase 1baseline benchmarking The aim of this phase was to set a baseline of all ten CNN architectures all operating under the same conditions with frozen convolutional weights and a newly trained classification head. This kind of arrangement offered an expedited and consistent place to compare architectural families prior to further optimization. We first conducted a preliminary experiment with all the ten architectures based on simple transfer learning where the convolutional part is frozen and only the classifier is trained. This gave us a quick and stable baseline to compare against. The performance across architectures is summarized in Table 5. Basab Nath, et al.: Comparative study of CNNs for brinjal disease detection8 Emirates Journal of Food and Agriculture The scores show that smaller, inception-style models (MobileNetV2, InceptionV3) worked better than deeper residual network or EfficientNetB0. Although the performance of ResNet50 and EfficientNetB0 were substandard, MobileNetV2 achieved an accuracy of more than 84% despite having been trained for short duration. These observations led us to choose to concentrate further tuning on the best families. From this simple comparison it can be observed that shallow and inception type models captured discriminative information without being guided by more than just frozen-base transfer learning constraints. This also indicates that MobileNet and Inception like architectures, which feature separate layers taking depthwise separable convolutions or multi-branch receptive fields, are particularly more suitable for the moderate variable nature of BrinjalFruitX. On the other hand, deeper residual/compound scaling models perform worse since they need much larger training sets to converge. The initial results of these bases were used to determine MobileNetV2, InceptionV3, and VGG16 as potentially the most promising architecture to consider tuning later. Based on these preliminary observations, we focused on the MobileNet, Inception and VGG architectures in our experiments. Phase 2 - hyperparameter tuning This phase designed to cautiously tweak the training dynamics of our best networks from Phase 1 MobileNetV2, InceptionV3 and VGG16 with a systematic approach that tried learning rates and batch sizes in order to find appropriate configuration that leads into stable convergence/ globalization. We conducted a controlled hyperparameter sweep over the best performing models from Phase 1: MobileNetV2, InceptionV3 and VGG16. We also investigated the effects of batch size (8, 16, 32) and learning rate (1*10−4 vs. 1*10−5). Results are shown in Table 6. Two general findings were robust to all models. First, configurations trained with a learning rate of 1*10−4 converged faster and at higher peak validation accuracy than those trained with 1*10−5, which often plateaued at a lower level. Second, smaller batch sizes (8 or 16) led to smoother learning dynamics and better generalization as compared to larger batch sizes (32), which tended to level off at lower accuracy. Critically, the favored setups yielded stable validation plateaus without departing from training curves, reflecting successful generalization and minimal overfitting under our augmentation policy and early stopping regime. In addition to these trends, architectures within each family showed distinct performance behaviours. MobileNetV2 gained a lot from batch size 8 which indicates that diversity in gradients stabilized the depthwise separable layers. InceptionV3 worked best with batch size 16, which is expected given the small overhead it has to compensate for its multi-branch nature. VGG16, on the other hand, was quite insensitive to the batch-size variation but started to show small performance gains as batch-size grew. These tuning results validated two crucial design decisions going into the next stage: choosing 1*10−4 as the default learning rate and opting for smaller batch sizes for the MobileNet and Inception families. All the architectures were re-assessed in the four training strategies mentioned above based on these optimized hyperparameters. Using these best settings, we moved to Stage 3 and compared the variants starting from all ten architectures in a carefully organized way for all four training packet strategy. Phase 3comprehensive strategy evaluation This last stage intended to make an “all-ten-architectures-off” benchmarking under the umbrella of four differing training methods (From Scratch, Transfer Learning, Fine-Tune and Full Monty) probing into how the selected learning strategy affects accuracy and generalization ability of models while keeping track on their computational budgets. Following these observations, in the last step we did an extensive benchmarking comparing task performance with respect to how each training strategy affects final accuracy, generalization and computational efficiency among the ten architectures. There are few observations that can be seen from these findings. First, the strategy selected was found to dramatically impact performance. The Full Monty methodology consistently obtained the best results among strategies for MobileNet, Inception and VGG families of architecture over that achieved by simple transfer learning. Second, performance was widely different across the architectural families: MobileNetV2 (Full Monty) Table 6. Effect of batch size and learning rate on selected models during Phase 2. Model Batch Size Acc @ LR = 1 × 10–4 (%) Acc @ LR = 1 × 10–5 (%) MobileNetV2 8 91.45 86.32 16 90.82 85.70 32 89.54 84.10 InceptionV3 8 88.67 82.93 16 89.92 84.27 32 87.54 81.65 VGG16 8 84.23 80.15 16 85.02 80.74 32 85.89 81.12 Table 5. Baseline performance of all CNN architectures under transfer learning during Phase 1. Model Test Accuracy (%) Test Loss Macro-F1 Weighted-F1 MobileNetV2 84.76 0.412 0.846 0.852 InceptionV3 82.45 0.435 0.821 0.829 VGG16 79.32 0.487 0.793 0.801 VGG19 78.94 0.493 0.789 0.798 InceptionResNetV2 77.56 0.521 0.775 0.783 ResNet50 52.08 0.912 0.510 0.523 ResNet152 54.32 0.876 0.529 0.542 EfficientNetB0 40.17 1.134 0.401 0.418 MobileNetV3Large 68.12 0.657 0.673 0.681 Emir. J. Food Agric ⋅ Volume 37 ⋅ 2025 9 Emirates Journal of Food and Agriculture obtained the best trade-off of accuracy and efficiency (97.98% mel-accuracy, macro-F1 = 0.9793) for short training times whereas InceptionV3 and VGG16 also yielded competitive accuracies but relying much higher computations. Third, several families of more complicated models (ResNets, EfficientNetB0) performed worse due to their architectural complexity and also because these models require large datasets to achieve a good score. The overall results are shown in Table 7, indicating test accuracy, test loss, macro and weighted-F1 scores and total training time. Final efficiency analysis and deployment suitability This phase measured the computational and deployment efficiency of the best performing architectures with a focus on the trade-off between prediction accuracy, number of parameters and inference latency required for practical applications at field scale. To complement accuracy-based comparisons, we also analyzed model efficiency in terms of parameter count, FLOPs, and inference time (Table 8). Lightweight architectures such as MobileNetV2 (3.4M params, 0.30B FLOPs) provided strong accuracy with a low computational footprint, while VGG variants and InceptionResNetV2 were substantially heavier. Table 8. Efficiency comparison of the evaluated architectures. Model Architecture Parameters (M) FLOPs (B) Inference Time (ms) MobileNetV2 3.4 0.30 3.2 MobileNetV3Large 5.4 0.22 3.5 EfficientNetB0 5.3 0.39 4.1 InceptionV3 23.8 5.7 8.9 InceptionResNetV2 55.9 13.2 15.7 VGG16 138.4 15.3 21.5 VGG19 143.7 19.6 24.8 ResNet50 25.6 4.1 7.6 ResNet152 60.2 11.6 18.3 Table 7. Comprehensive performance evaluation of all models and training strategies. Model Architecture Strategy Name Batch Size Test Accuracy (%) Test Loss Macro Avg F1 Weighted Avg F1 Training Time (HH:MM:SS) MobileNetV2 Full Monty 8 97.98 0.0707 0.9793 0.9799 00:18:42 Fine-Tune 8 94.35 0.1851 0.9425 0.9438 00:34:12 Transfer Learning 8 94.76 0.1836 0.9466 0.9478 00:07:17 From Scratch 8 58.47 0.6789 0.3690 0.4314 00:12:21 InceptionV3 Full Monty 897.58 0.1027 0.9750 0.9758 00:55:39 Fine-Tune 8 89.92 0.2717 0.8966 0.8994 00:54:13 Transfer Learning 8 89.52 0.2724 0.8926 0.8954 00:19:32 From Scratch 8 93.55 0.2184 0.9328 0.9350 00:44:49 VGG16 Full Monty 16 97.58 0.1006 0.9751 0.9758 00:45:12 Fine-Tune 16 85.89 0.4426 0.8560 0.8594 00:24:44 Transfer Learning 16 87.10 0.4375 0.8688 0.8717 00:23:49 From Scratch 16 89.92 0.2788 0.8969 0.8995 01:09:54 InceptionResNetV2 Full Monty 16 96.77 0.0847 0.9668 0.9677 01:55:43 Fine-Tune 16 83.87 0.3464 0.8383 0.8397 01:40:33 Transfer Learning 16 84.27 0.3423 0.8420 0.8438 01:05:02 From Scratch 16 96.77 0.1783 0.9668 0.9677 02:52:20 VGG19 Full Monty 16 95.97 0.1033 0.9582 0.9596 00:55:10 Fine-Tune 16 85.89 0.4311 0.8516 0.8571 00:30:58 Transfer Learning 16 87.50 0.4248 0.8680 0.8732 00:30:39 From Scratch 16 89.52 0.2787 0.8929 0.8955 01:23:02 ResNet50 Full Monty 16 91.13 0.2452 0.9065 0.9101 01:14:26 Fine-Tune 16 68.15 0.6296 0.6434 0.6631 00:39:31 Transfer Learning 16 66.53 0.6348 0.6438 0.6586 00:23:34 From Scratch 16 88.71 0.3176 0.8831 0.8867 00:54:34 EfficientNetB0 From Scratch 16 88.71 0.2548 0.8838 0.8871 00:56:22 Fine-Tune 16 58.47 0.6829 0.3690 0.4314 00:11:46 Transfer Learning 16 58.47 0.6851 0.3690 0.4314 00:08:58 Full Monty 16 64.11 0.6333 0.6300 0.6191 00:16:13 MobileNetV3Large From Scratch 8 58.47 0.6787 0.3690 0.4314 00:26:39 Fine-Tune 8 70.97 0.6451 0.6560 0.6790 00:21:25 Transfer Learning 8 72.98 0.6480 0.6919 0.7102 00:07:39 Full Monty 8 70.56 0.6484 0.6383 0.6647 00:11:37 ResNet152 From Scratch 16 89.52 0.2791 0.8907 0.8945 02:54:34 Fine-Tune 16 71.37 0.6304 0.6897 0.7043 01:51:00 Transfer Learning 16 71.77 0.6271 0.6976 0.7108 01:01:02 Full Monty 16 87.10 0.2862 0.8655 0.8701 03:45:11