A Resource-Efficient YOLO Framework via Discrete Wavelet Transform-Based Image Compression for Plant Disease Detection

2.1 Lossy compression Discrete Wavelet Transform

The wavelet transform is a signal-processing method that analyzes data based on frequency differences and determines the locations of those frequencies. Applications include watermarking [27], health detection [8], sensing [28], and solving differential equations [29]. The approach involves two main stages: the DWT separates signal components into low- and high-frequency components, and the Inverse Discrete Wavelet Transform (IDWT) reconstructs the original signal.

The DWT stage is typically used to reduce image size by quantizing high-frequency components. This process enhances the efficiency of data storage and transmission. The IDWT stage restores the transformed image components to their original form. For compressed images, DWT, IDWT, and dequantization enable image reconstruction. This maintains image quality close to the original. Thus, the wavelet transform plays a significant role in image processing. It is particularly useful in data compression and image quality restoration.

Originally developed to analyze non-stationary signals via multi-resolution analysis, the wavelet transform allows observation of frequency components at several resolution levels. Quantization of DWT coefficients introduces limited information loss; however, when combined with appropriate encoding schemes, high reconstruction fidelity can be preserved. For storage and transmission efficiency, studies use canonical Huffman coding or combine wavelet filtering with Huffman coding [6, 30].

In the continuous wavelet transform, the scale (a) and translation (b) parameters vary continuously, resulting in high computational complexity. The DWT addresses this by restricting a and b to discrete multiples of two, enabling more efficient analysis using specific values. DWT processes large data sets faster and more effectively. Eq. (1) defines the mathematical formulation of the wavelet function based on discrete scale and translation parameters [31].

$\varphi_{a, b}(t)=\frac{1}{\sqrt{|a|}} \varphi\left(\frac{t-b}{a}\right)$    (1)

where, $\varphi_{a, b}(t)$ denotes the wavelet coefficients at scale $a$ and shift $a$, with $x(t)$ as the input signal and $\varphi(t)$ as the wavelet function. Parameter $a$ governs compression/dilation, while $b$ determines the position in the time domain. Expanding Eq. (1) by substituting several parameters, $a=2^{\mathrm{r}}$ and $b=k .2^{\mathrm{r}}$, then produces Eq. (2).

$\varphi_{r, k}(t)=\frac{1}{\sqrt{\left|2^r\right|}} \varphi\left(\frac{t-k \cdot 2^r}{2^r}\right)$    (2)

Description $\varphi_{{{r,k}}}(t)$ is a discrete wavelet function with $r$ as the scale and $k$ as the translation. Substitution of Eq. (2) into the continuous wavelet transform produces the discrete form, as shown in Eq. (3).

$D W T_{r, k}=\frac{1}{\sqrt{\left|2^r\right|}} \int_{-\infty}^{+\infty} x(t) \varphi\left(\frac{t-k \cdot 2^r}{2^r}\right)$    (3)

The calculation of the discrete wavelet coefficients $D W T_{r_, k}$ is performed by applying a discrete scale parameter $2^r$ and a translation $k .2^r$ to the signal $x(t)$ through the wavelet function $\varphi(t)$. This formulation represents the result of substituting the continuous wavelet transform into a discrete form, designed to optimize computational efficiency for large-scale datasets.

2.2 YOLO object detection

Object detection is a key area of computer vision, aiming to recognize and locate objects in images or videos. The YOLO algorithm was developed as a one-stage detector that formulates object detection as a single regression, enabling it to predict object class and bounding box coordinates directly in a single process. This approach makes YOLO faster and more efficient than two-stage methods such as R-CNN and its derivatives, making it widely used in various applications requiring real-time detection [32, 33].

YOLOv8 is one of the latest versions, introduced by Ultralytics in 2023. YOLOv8 brings several innovations. It uses a C2f module in the backbone for more efficient feature extraction. It combines a Feature Pyramid Network (FPN) and a Path Aggregation Network (PAN) in the neck for better multi-scale detection. It also implements an anchor-free head for direct object prediction, removing the need for anchor boxes. These architectural improvements enhance detection accuracy and computational efficiency, making YOLOv8 suitable for real-time vision applications. Thus, it is relevant for fields from intelligent surveillance to industrial manufacturing [34, 35].

Image compression is a crucial process in digital image processing that reduces data size without significantly compromising visual quality. One widely used method is the DWT due to its ability to represent images in the frequency domain with multi-resolution analysis. Through the decomposition process, DWT divides the image into low-frequency components that retain the main information and high-frequency components that contain fine details, thus enabling efficient data reduction. Table 3 presents sample image compression results, including PSNR and MSE values.

The results of the dataset image compression in Table 3 show that the image compression results produce different values. The dataset image sampling is the result of camera acquisition with different specifications, as shown in Table 1. The largest difference in compression results is in data 9, which is the result of acquisition from the FIMI X8 Mini drone camera. Based on the image compression results, the next step is to detect objects using the YOLO object detection algorithm. The dataset, in the form of plant images, will be classified into three classes: banana, mango, and guava. In this study, the dataset labeling utilizes two software tools, namely the CVAT and LabelImg.

Table 4 shows the resulting images from image compression (before and after compression). The images before and after compression show very little difference. Visually, the original image has slightly sharper details in the texture of the banana leaf and stem, with smoother color transitions. After compression, some fine details appear slightly reduced and object edges feel softer, but the main plant structure is maintained. The quantitative analysis results also support this observation, showing minimal pixel differences, while the very high SSIM value indicates that the structures of the two images are almost identical. Overall, the compression process does not destroy important information in the image and is still very suitable for use in object detection processes such as YOLO, because the main shape and pattern of the plant are well preserved.

Table 3. Dataset image compression results

Note: MSE = mean squared error; PSNR = peak signal-to-noise ratio.

Table 4. Comparison of images before and after image compression

Before	After

The results of the classification model evaluation (Figure 6), based on the confusion matrix, show an overall accuracy of 25.6%, with macro-average precision, recall, and F1-score values of 25.4%, 24.5%, and 24.2%, respectively. The main errors occurred in the model's tendency to classify fruit objects as background objects, as well as overlapping predictions between fruit classes, particularly between mango and guava. These findings indicate the need for improvements to the training data and model architecture to more accurately distinguish the visual characteristics of each class.

Figure 6. Confusion matrix

Figure 7, a recall-confidence graph, shows that recall decreases as the confidence threshold increases. The guava class performed best, with a relatively higher recall compared to the banana and mango classes, whose recalls declined rapidly at higher thresholds. Overall, the model achieved a recall of 0.69 at a threshold of 0.0; however, performance decreased significantly as the confidence threshold was increased. This confirms the trade-off between recall and prediction confidence.

Figure 7. Recall graph

Figure 8, a precision-confidence graph, shows that precision values increase as the confidence threshold increases. The guava class has the highest precision (> 0.8), followed by mango, which shows moderate performance, while banana exhibits very low precision. Overall, the model achieved a precision of 1.00 at a threshold of 0.961, although this potentially reduces recall. These results confirm the trade-off between precision and recall in model performance.

Figure 9 shows the precision-recall curve, showing that the guava class performed best (AP = 0.416), followed by mango (AP = 0.340), and banana performed the lowest (AP = 0.086). The average mAP@0.5 value of 0.281 indicates that the model's performance is still low and requires improvement through improvements to the dataset and model architecture.

Figure 8. Precision value

Figure 9. Precision-recall graph

Figure 10. F1 graph

Figure 10 is an F1-Confidence graph showing the variation in F1-score values against the confidence threshold. In general, the model achieved the highest average F1-score of 0.34 at a threshold of 0.176. The guava and mango classes performed relatively better than the banana class, with their respective F1 peaks approaching 0.5. Meanwhile, the banana class tended to have low stability and the lowest F1-score along the curve. These results indicate that model performance is more consistent across certain classes.