GLAD: Advanced Attention Mechanism-Based Model for Grape Leaf Disease Detection

Grapes are an economically valuable and widely consumed fruit. Grape leaves are highly vulnerable to infections from bacteria, viruses, and fungi. The quality and quantity of grapes can be diminished if these diseases spread throughout the plant. Diseases on grape leaves used to be identified manually using standard phytopathology techniques. It's a lot of work and takes a long time, though. As more and more acreage is devoted to grape cultivation, the accuracy of manual identification methods declines. For the sake of grape production's continued growth, automatic identification of leaf diseases is essential. Automated identification techniques are more efficient, precise, and trustworthy than their manual counterparts. Also, unlike manual methods, they may be utilized to keep tabs on broad stretches of land.

Maintaining healthy crops, decreasing crop losses, and using as little pesticides as possible are all crucial in agriculture [1]. The key to successful disease prevention and early diagnosis is vigilance. Powdery mildew, brown blotch, and anthracnose are just a few of the diseases that can affect grape vines. The quantity and quality of fruit produced may be drastically affected by these illnesses. Detecting grape diseases via tried-and-true approaches like producers' anecdotes or experts' advice can be time-consuming, inefficient, and behind the times. Grape leaf photos are utilized to detect, identify, and direct disease prevention efforts [2] to overcome these constraints. The leaves of sick grapes tend to be discolored and distorted.

Classifying and identifying crop diseases using traditional methods [3-5] requires laborious manual feature selection that is highly context dependent. However, deep learning's introduction of convolutional neural networks (CNNs) has radically altered the process of identifying agricultural diseases. A lot of work has been done to categorize agricultural illnesses using various methods [6-12]. By extracting high-dimensional features from photos of the sick leaves, CNNs can accurately diagnose crop leaf illnesses even in the presence of complicated backdrops.

Face recognition [13], navigation systems [14], obstacle identification [15], pedestrian detection [16], abnormal activity recognition [16], monitoring physical activity [17, 18], fruit detection [19], and weed detection [20] are just a few examples of the many uses for CNNs. In order to help farmers reduce crop failures and maximize yields, these technologies can be utilized to perform automatic identification of crop leaf diseases. Scientists from China and other nations have been investigating object detection algorithms to create crop disease detection models. They have tried out various models using tomato disease datasets, including Faster R-CNN and Single Shot Multibox Detector. The most encouraging results in disease identification have been achieved by combining faster R-CNN with VGG16. Researchers have successfully used Faster R-CNN on time-series images of grape leaves for dynamic identification of grape leaf diseases. Using an improved Faster R-CNN model, the authors [20] were able to get good disease detection results in bitter gourd leaves. The authors [21] also used an internal dataset to train the SSD (Single Shot Multibox Detector) model, which led to an 83.90% overall accuracy in detecting agricultural illnesses.

Several methods, such as an improved model based on MobileNetv2 and YOLOv3 [22], have been proposed by researchers to identify agricultural illnesses. This model's many strengths include its small memory footprint, high detection accuracy, and lightning-fast identification times. However, CNNs' limited perception is a significant limitation. Because of convolutional computation, CNNs often only extract characteristics from nearby regions of an image [23].

We propose GLAD, a deep learning model for grape leaf disease detection, to overcome this shortcoming. With GLAD, CNN-based object detection is improved over YOLOv5 [24]. GLAD can gain a comprehensive overview of the image by including the self-attention mechanism. This allows GLAD to ignore irrelevant noise and zero in on relevant regions, leading to more precise detections. The GLAD model is an advancement on the YOLOv5 deep learning model for detecting diseases in grape leaves. It facilitates the model's acquisition of more contextual information by creating a global perception eld [25]. Several novel elements are included in the GLAD model. They are

• As a result of BIFPN [26], grape leaf diseases are enhanced on multiscale features, and features become more prominent.
• The Shuffle Attention mechanism [27] enables the GLAD model to place greater emphasis on grape leaf diseases during the detection process.
• The ASFF [28] detection head automatically filters out irrelevant information and suppresses interference from complex backgrounds when detecting grape leaf diseases.

Through transfer learning, the GLAD model can be accelerated in situations with limited sample sizes with an increase in accuracy and robustness.

3.1 Image acquisition

Detection and classification of grape diseases are made easier using the plant village dataset. Four thousand six hundred and twenty-two images show grape leaves with many symptoms, including black rot, esca measles, leaf spot, and healthy leaves. The dataset is imbalanced, with Black Rot being the largest category (1,180 images). Esca measles is the most common classification, followed by Leaf spot. The dataset also contains 423 images of healthy grape leaves for reference. The varying quantities of images indicate the prevalence of certain grape leaf diseases. Esca measles accounts for approximately 34% of the dataset, while Black Rot and Leaf spot represent 29% and 26%, respectively. Healthy images make up about 10.4% of the collection. The plant village dataset provides researchers with a valuable resource to train and evaluate grape disease detection models. It can be used to develop more accurate algorithms for automated diagnosis and classification of grape leaf conditions.

1.png

Figure 1. Sample images from dataset

Table 1. Prior works on grape leaf disease detection

2.png

Taking all of these things into account, the researchers chose YOLOv5 as the basic framework for spotting diseases on grape leaves. They then made some new changes and developed the GLAD model, shown in Figure 2.

3.3.2 Self-attention system of transformer

Images of grape leaves can show how diseases are spread in different ways. Some diseases, like leaf blight, only affect small parts of a leaf and stay in one place. Detection depends more on local knowledge and high-level features in these situations. Other diseases, like black rot, are all over the leaf and can only be found by looking at the whole plant. It is important to get global semantic knowledge to improve the network's ability to figure out where it is. Most of YOLOv5's basic structure is based on convolutional neural networks (CNNs). CNN can only get information about long-distance interactions if they focus on how images next to each other are related. In addition, they need to make long-range semantic links better, and that's a problem for networks that want to ignore noise all over the feature map to focus on specific areas of interest. There is a strong correlation between the actual perception elds of extracted features and their theoretical counterparts, as they are significantly smaller in reality. This shows that convolutional operations alone aren't enough to determine how local picture features depend on things far away.

To get around this problem, the researchers developed GLAD, a new model that includes the self-attention process. Figure 3 shows the self-attention mechanism, which lets the model take in global semantic knowledge and improve its ability to and its way around.

3.png

Figure 3. Self attention system for GLAD model

Self-attention processes have been suggested to get around the fact that CNNs are inherently local. The Transformer is a self-attention device that has been shown to work well in tasks like processing natural language and computer vision. The Transformer works by paying attention to different parts of the input series and learning how they are related. This means that the Transformer can pick up long-distance relationships that would be hard for CNNs to pick up.

Adopting visual Transformer networks shows how important it is to set up remote dependencies if you want to make big changes to how computer vision jobs are done. In short, adding self-attention mechanisms, especially the Transformer, gets around the fact that CNNs can't capture long-range dependencies and lets the network use global semantic information successfully. This makes it easier for the network to nd and pinpoint tea diseases. The attention can be calculated as:

$S_a={attention}\left(Q_v, K_v, V_v\right)={Softmax}\left(\frac{Q K^T}{\sqrt{d_k}}\right) V$        (1)

where, S_a is the self-attention mechanism, Q_v is the query vector, K_v is the key vector, V_v is the value vector.

3.3.3 Enhanced object detection through multiscale feature fusion with BIFPN

Within the same disease category, our collection of grape leaf diseases revealed a broad variation of sizes and forms. This necessitates incorporating traits from different scales. To make YOLOv5's multiscale feature fusion network more effective, we swapped out the original PAFPN for BiFPN. Multiscale feature fusion is a module that is part of the BiFPN plugin. This updated PAFPN for YOLOv5 incorporates residual connections into the actual network topology and is based on EfficientDet. Nodes that contribute less to feature fusion are removed using cross-connections in PAFPN, while input and output nodes of the same scale are connected using jump connections. In contrast to PAFPN, BiFPN distributes feature information uniformly across multiple scales using weights.

Our GLAD model efficiently fuses multiscale aspects of grape leaf diseases by adhering to the concepts of BiFPN. As a result, the GLAD model's capacity for representing features is improved while the number of parameters is decreased. Figure 4 depicts the internal organisation of BIFPN.

4.png

Figure 4. Structure of BIFPN

3.3.4 Attention mechanisms

Shuffle Attention: Detecting small target grape leaf diseases can be improved using the Shuffle Attention (SA) mechanism. SA combines spatial attention and channel attention mechanisms within blocks, which allows for their effective integration. Spatial attention captures the spatial dependence between pixels, while channel attention captures the dependence between channels. Using both types of attention mechanisms simultaneously can yield better results but requires increased computational resources. SA addresses this issue by incorporating the Channel Shuffle operation, which allows for efficient computation of both spatial and channel attention.

First, SA incorporates global information and generates a channel feature using Global Average Pooling as s e R^C/2Gx1x1. Z_k1. It is incorporated in Eq. (2).

$\mathrm{S}=\mathrm{F}_{\mathrm{gp}}\left(\mathrm{Z}_{\mathrm{k} 1}\right)=\frac{1}{A \times B}=\sum_{m=1}^A \sum_{n=1}^B\left(Z_{k 1}\right)(\mathrm{m}, \mathrm{n})$         (2)

An adaptive selection module based on this channel feature is then applied, along with an activation function based on sigmoid geometry, in order to produce a precise feature that provides guidance for adaptive selection. It is incorporated in Eq. (3).

$Z_{k 1}^{\prime}=\sigma\left(F_c(s)\right) \cdot Z_{k 1}=\sigma\left(V_1 s+b_1\right) Z_{k 1}$     (3)

The SA mechanism then uses group norm (GN) to normalize the spatial feature map Z_k2. This normalized feature map is then passed through Fc(·) to enhance its representation. Spatial attention can be calculated using Eq. (4) as the final output.

$\left.Z_{k 2}^{\prime}=\sigma\left(V_2 \cdot G N Z_{k 2}\right)+b_2\right) \cdot Z_{k 2}$        (4)

The SA mechanism can be used to improve the detection of small target grape leaf diseases by selectively focusing on the characteristics of these diseases. This is done using spatial and channel attention mechanisms to capture the relevant information from the input image. The SA mechanism is a promising new approach for improving the accuracy of object detection algorithms.

The initial convolutional neural network (CNN) design of YOLOv5 included shuffle attention. Shuffle attention improves the model's capacity to convey the semantics of grape leaf disease traits by efficiently merging channel and spatial attention processes. One benefit of the suggested GLAD model is its ability to be partitioned and parallelized, which allows for the extraction of attention regions and the focus on the unique features of grape leaf diseases.

Convolutional Block Attention Module: Convolutional neural networks (CNNs) can be more precise with the help of an attention mechanism called the Convolutional Block Attention Module (CBAM). CBAM analyzes the output of a CNN, and the model's ability to classify images is enhanced by emphasizing only the most relevant elements. By zeroing in on the most pertinent aspects of an input image, CBAM can improve the detection of small-target grape leaf diseases. A combination of spatial and channel attention techniques is used to pick out the crucial details in the input image and bring them to the fore.

CBAM's spatial attention mechanism is useful for pinpointing key regions of interest within an input image. To create an attention map, we combine the results from the CNN's convolutional layers and apply a sigmoid activation function to the combined data. By highlighting the most crucial regions of the input image, the attention map directs the CNN's processing power where it is most needed.

CBAM's channel attention mechanism aids in selecting the most relevant channels from the given input image. A sigmoid activation function is applied to the combined output of the CNN's convolutional layers across the channel dimension to create an attention map. By highlighting the most crucial channels in the input image, the attention map directs the CNN's processing power where it is most needed.

CBAM can aid in the accuracy of CNNs by selecting and focusing on the most significant elements in the input image through a combination of spatial and channel attention techniques. This is especially useful for jobs that require discriminating between items or classes, such as object detection. By including it in the model's feature extraction layers, YOLOv5 now benefits from the CBAM attention module.

Efficient Channel Attention: ECA (Efficient Channel Attention) is a channel gating method that uses a feature map to learn the relevance of each channel. ECA is a superior attention mechanism because it uses a more effective method than SE-Net's global average pooling layer. This can be useful for reducing the computational complexity of the model without sacrificing accuracy, which is especially relevant for disease detection in grape leaves. First, a weight is determined for each channel in the feature map, which is used by the channel gating mechanism in ECA. The feature map is gated using these weights so only channels with large values can proceed to the next layer. By zeroing in on the most crucial channels in the feature map, ECA is able to boost the precision of the model. SE-Net's global average pooling layer assigns a weight to each channel in the feature map. However, unlike ECA's channel gating technique, the global intermediate pooling layer averages all the pixel values in each channel.

ECA's potential to reduce the model's computational complexity makes it a promising tool for disease detection in grape leaves. As disease identification in grape leaves is a computationally intensive task, this is significant as it can lead to better model speed and performance.

Multi-Channel Attention (MCA): It combines a single-stage attention mechanism with multiple attention heads to learn the significance of distinct channels on a feature map. As a result, MCA can pick up on subtler interdependencies between the channels, which can prove helpful in situations like disease detection in grape leaves. For MCA to function, the feature map must rst be segmented into separate channels. Then, individual "attention heads" deal with each channel individually. Each channel's attention head figures out how significant its channel is by gauging it against the others. With several attention heads, MCA can learn intricate interdependencies between channels. Complex dependency learning might be helpful for problems like disease detection in grape leaves, where distinguishing between similar diseases can be difficult. For instance, the symptoms of two distinct grape leaf diseases may be identical, yet their effects on the leaves may differ. MCA can discern these subtle distinctions by learning the interdependencies between the channels in the feature map.

3.3.5 Adaptive spatial feature fusion

Grape leaf disease detection is difficult because of the diverse context and changing disease severity. To solve these problems, we suggest a new detection method called adaptive spatial feature fusion (ASFF). Grape leaf disease diagnosis is greatly aided by ASFF's ability to automatically alter out the noise and minimize interference from complicated backdrops. It also makes it easier to combine data from multiple levels of illness analysis.

We swapped over YOLOv5's old detecting head for a new ASFF one. Adjusting the fusion ratio between feature layers is a key function of the ASFF detection head. During multiscale feature fusion, ASFF decreases inference overhead by altering spatially connecting information and improves the invariance of the feature ratio. As a result, these targets for grape leaf diseases can be detected more easily. Figure 5 depicts the internal organization of ASFF.

The YOLOv5 network's "neck" produces results labelled "Level1, Level 2, and Level 3." For illustration, think about ASFF-3. After fusion, ASFF-3's input is derived from a weighted total of the inputs at Levels 1, 2, and 3. Multiplication and adding of masses characterize this fusion process, as seen in Eq. (5).

b'_ij=a_ij’ * a_ij^1àl +b_ij’* a_ij^2àl +c_ij’ * a_ij^3àl         (5)

5.png

Figure 5. Structure of ASFF