Tomato, as a main cultivated vegetable with a worldwide production of over 170 million tons annually, draws huge attention on its disease prevention. Numerous types of tomato diseases that target the tomato's leaf at an alarming rate have been discovered. Though traditional method requires high human involvement and people are seeking autonomous method, the early symptoms of tomato disease which can be observed on leaves are too tiny for existing lightweight detectors. By improving the backbone, we have proposed a lightweight object detection model that improves feature extraction, especially for tiny objects, compare to the original YOLOv8. We have benchmarked the original YOLOv8 and our improved model at diferent input image sizes and model scales. Experiments on the tomato leaf disease dataset have shown that the best accuracy at mAP50 is 1.8% higher than YOLOv8l. Additionally, our improved model has exhibited a 74% reduction in parameter size.
one of the most popular domains within computer-vision science. Currently, depending on the
method of implementation detection algorithms can be categorized into two groups. Two-stage
models, including SPPNet[
In this research, the aim was to improve the accuracy of detection on the tiny symptoms of tomato disease which can be observed on leaves. To address this challenge, we have adapted the channel attention mechanism and compound scaling method to construct our backbone network for better feature extraction ability. To achieve mobility as well as high performance of our proposed model, we have searched our whole network with multiple combinations for the best model structure. Finally, our proposed model has outperformed the original YOLOv8 on datasets with classes that have indistinct appearance without significantly increasing its number of parameters or inference time.
Overall, our main contributions can be summarized below: • An improved model for tiny objects has been proposed by combining the MBConv module and YOLOv8 head. • For usability on mobile devices, a series of models with diferent sizes have been searched and eventually a lightweight network has been proposed.
The remainder of this article is organized as follows. Section 2 reviews the previous agricultural detection and related YOLO models. Details of our proposed method are introduced in Section 3. Section 4 presents the experiments and the dataset. Finally, Section 5 concludes the paper.
Object detection aims to classify certain objects from defined categories contained in the dataset
and then predict the position of the object. The challenges in object detection can be basically
categorized into increasing the inference speed and enhancing the accuracy of classification, as
well as predicting object location. Benefiting from the development of deep learning, improved
accuracy and detection speed can now be achieved with state-of-the-art deep learning algorithms.
YOLO models have been used for many purposes due to their superior performance, especially
for detection tasks in videos[
To achieve maturity detection, positioning, and harvesting in crop production, as well as disease, pest, and climate hazard prevention during the growth process, object detection technology is being introduced into the fields to enhance agricultural productivity. Taking tomato cultivation as an example, classification networks can be used to categorize potential diseases on tomatoes.
TM et al.[
To achieve real-time object detection and automation in industrial production or services, the YOLO series continues to play a vital role due to its high speed, high accuracy, lightweight and locally deploy-able features.
Egi et al. [
In this study, we have utilized network structure extracted from Eficient-Net[
YOLOv8 consists of a backbone network and a head network. With scaling factors N/S/M/L/X, YOLOv8 provides five models of diferent sizes, each possessing unique characteristics in terms of the number of parameters and required computational resources. These diferences result in significantly diferent detection performance and inference times.
The backbone network utilizes convolution operations to extract features. In this process, the
feature map gradually downgrades its size while increasing its number of channels. As a result,
the backbone network can retain and extract important features passed to the head network for
further processing. The head network is mainly designed to fuse features from shallow layers
representing geometry and features from deep layers representing semantics. Compared to
its predecessor, YOLOv5[
In practice, we have observed that the performance of classification models can be improved
with appropriate increments in depth, width (number of channels), or resolution, but the positive
gain decreases as the model becomes larger. To address this problem, Network Architecture
Search (NAS)[
= ,,
( (, , )) .. (, , ) =
ˆ · ^ (<· ^,· ^ ,· ^>) ⨀︀ =1... ( ) ≤ _
( ) ≤ _ where r,w,d are coeficients for resolution, width and scaling network depth: ˆ , ˆ, ˆ, ˆ , ˆ are predefined parameters in the baseline network Eficient-Net B0. To avoid tedious manual tuning on balancing network width and depth, a new compound scaling method has been proposed which use a compound coeficient to uniformly scales network depth, width and resolution in a principled way: (1) (2) (3) (4) (5) PP44//1166 P/8 3
Original Eficient-Net has been trained and tested on classification tasks. In this study, we aimed to leverage its strong ability on feature extraction and lightweight design so as to achieve better performance on tiny object detection without significantly increasing parameters and FLOPs compared to YOLO series. As shown in Figure 1, the model consists of two parts, Backbone and Head.
Feature extraction is vital for object detection, thus, we aimed to improve the backbone. As shown on the left of Figure 1, the MBConv module has been repeated to form a feature pyramid network. Inside the MBConv module, the squeeze and excitation layer provides channel attention, and depth-wise separable convolution significantly decreases model size. By adjusting the coeficients of depth, width, and resolution of the backbone, models of diferent sizes for specific applications can be conveniently constructed.
As shown on the right of Figure 1, we have used two components: the Feature Pyramid Network (FPN) [28] and the Path Aggregation Network (PAN) [29]. FPN first extracts feature maps from the backbone to construct a feature pyramid. It merges feature maps from diferent levels through up-sampling and coarser-grained feature maps, enabling the fusion of diferent feature levels. The C2f module is used to keep the model light-weighted and enable richer gradient information. Furthermore, the decoupled detect module separates classification and regression bounding boxes with two independent loss functions. For classification loss, Binary Cross Entropy (BCE) Loss is used to calculate the loss for each category, and for bounding box loss, Complete Intersection over Union (CIoU) is used to calculate the diference between the bounding boxes of objects and predictions. Distribution Focal Loss (DFL) is then used to address the extreme imbalance between foreground and background classes.
In this section, comprehensive experiments have been conducted using models with backbones and heads of diferent sizes on the tomato leaf disease dataset, demonstrating the efectiveness of the proposed network.
We have used the public Tomato Leaf Disease dataset[30] on roboflow, which consists of 7 categories of tomato leaf diseases: Bacterial Spot, Early Blight, Healthy, Late blight, Leaf Mold, Target Spot, Black Spot. Image augmentation has been applied to each photo to address the deficiency in data variety.
The Tomato Leaf Disease dataset has 737 images in total, 645 images for training set, 61 images for validation set, and 31 images for testing. As shown in the Figure 2, two main challenges can be found: 1. Tiny objects. Symptoms of tomato leaf diseases can be very small, these minuscule symptoms are even dificult for human eyes to capture. As shown in the Figure 2(b), the size of objects are mostly below 10% of the image size. 2. Multiple symptoms. It is not uncommon to find multiple symptoms on the same tomato leaf, and the appearance of diferent symptoms can be similar. This similarity makes distinguishing between diferent diseases a significant challenge.
All experiments have been performed on a platform with an NVIDIA GeForce RTX 4090 graphics card. The deep learning framework used was PyTorch 2.2.2 with CUDA 12.1. The proposed method in this study has been implemented based on Ultralytics 8.1.47, an innovative framework developed by the company that proposed YOLOv5 and YOLOv8.
The epochs and batch size have been set to 400 and 16, respectively, with patience set to 100. The IoU has been set to 0.7 for non-maximum suppression. All models have been trained using (a) Label distribution (b) Distribution of object sizes the AdamW optimizer, with an initial and final learning rate of 0.01, momentum of 0.937, and weight decay of 0.0005. The learning rate has adopted the warm-up cosine annealing algorithm. The learning rate of the model gradually increases within the first 3 epochs. In the experiment, nearly all models triggered early stopping around epoch 400, which means suficient training has been conducted on the whole dataset.
Typical metrics used for object detection tasks have been used to evaluate models in this study, including Precision, Recall, mAP50, and mAP50-95. The definitions of these metrics are as follows: = =
= =
+
+ ∑︀ (+1 − ) (+1) =1
1 ∑︀ (10) (11) (12) (13) (14)
The Average Precision (AP) of all classes is the area of the region below the precision-recall curve. represents the recall of the th value, and (+1) represents the highest precision value in the range to +1. The mAP is calculated by averaging the AP of each class in the dataset. mAP50 is obtained by averaging the AP (IoU = 0.5) of all classes, and mAP50-95 is obtained by averaging the mAPs at diferent IoUs between 0.5 and 0.95.
The comparison model for the original YOLOv8 in this study was reproduced using the Ultralytics framework provided by the author. All models shared the same configuration. To visually demonstrate the efect of our improved method, six images from the tomato leaf disease dataset were used for testing, and the results are shown in Figure 3. The results show that the bounding boxes of all objects have been efectively detected. In the six images, most objects are tiny, occupying an area only 1/100th of the whole image, and the diferences in the appearance of each class are indistinct, resulting in classification dificulties.
To address this challenge, our improved network has adopted the MBConv module with a Squeeze and Excitation (SE) layer. The channel attention mechanism helped the backbone focus on efective feature maps during training, improving feature fusion in deeper networks. This has been resulted in better feature extraction.
The results of the original YOLOv8 with scale parameter n on the tomato leaf disease dataset have been depicted in Table 1. To investigate the efectiveness of diferent backbones with diferent scales, experiments on each set of parameters have been evaluated. Though the depth, width, and resolution should comply with the compound scaling method, the model performance on object detection benefits from higher resolution. In Table 1 and 3, it was evident that the models trained with low resolution result in worse performance than their counterparts trained with higher resolution, thus the comparison has been conducted using an input image size of 640x640.
In terms of the number of parameters, from the original YOLO series in Tables3, it was evident that increasing the number of parameters in the model did not necessarily improve the detection ability of the model. Compared to the original YOLOv8 with scale factor l which has the best performance on the task, our improved network with B2 coeficients and YOLOv8 head in m size has achieved the best performance on the evaluation metrics mAP50 with only 26% of the parameters but a 1.8% improvement. Therefore, the improvement in feature extraction ability of the backbone has proven its parameter efectiveness in enhancing object detection performance.
In this study, we have proposed our improved network designed for detecting tomato leaf diseases. Our network has improved the feature extraction ability of the backbone by adopting the module and structure from Eficient-Net. The experiments conducted on our improved network and comparison networks have shown that our network, which consists of an EficientNet backbone constructed with B2 size coeficients and YOLOv8 head constructed with m size coeficients, achieves the best mAP50 and mAP50-95, outperforming the original YOLOv8 models and proving the efectiveness of our model. In summary, our improved network with EficientNet backbone has satisfactory feature extraction capabilities, especially enhancing the object detection performance on tiny objects. Codes are available at https://github.com/radiuson/EfiYOLOv8. tional conference on computer vision, 2019, pp. 1314–1324. [26] J. Hu, L. Shen, S. Albanie, G. Sun, E. Wu, Squeeze-and-excitation networks, 2019.
arXiv:1709.01507. [27] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, L. Fei-Fei, Imagenet: A large-scale hierarchical image database, in: 2009 IEEE conference on computer vision and pattern recognition, Ieee, 2009, pp. 248–255. [28] T.-Y. Lin, P. Dollár, R. Girshick, K. He, B. Hariharan, S. Belongie, Feature pyramid networks for object detection, in: Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 2117–2125. [29] S. Liu, L. Qi, H. Qin, J. Shi, J. Jia, Path aggregation network for instance segmentation, in: Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 8759–8768. [30] S. A. University, Tomato leaf diseases detect dataset, 2024. URL: https://universe.roboflow. com/sylhet-agricultural-university/tomato-leaf-diseases-detect.