Application of Image Enhancement and Object Detection Technologies in Virtual Teaching Systems for Vocational Skills Training

Aiming at the problems commonly existing in vocational skills virtual teaching scenarios, such as high noise, low contrast, and inconspicuous color information in images, which are not conducive to target recognition and localization, this study proposes an image enhancement and object detection method that takes low-light teaching images as input. Firstly, the images are processed through an improved lightweight image enhancement network. This network utilizes a lightweight convolutional neural network to predict enhancement parameters and adopts a no-reference loss function to drive the learning process, without relying on annotated or paired image datasets. It can specifically improve the clarity, contrast, and color expressiveness of vocational skills teaching images, outputting enhanced version images suitable for target detection. Subsequently, the enhanced high-quality images are input into a lightweight improved YOLOv8n object detection network. By optimizing the model structure and parameters, the model can realize rapid and accurate detection and localization of specific targets such as tools, equipment, and operation parts in teaching images. This method, through a serial architecture of "enhancement first, then detection", effectively solves the adverse impact of image quality defects in vocational skills virtual teaching images on target detection, providing key technical support for intelligent interaction and automatic assessment functions of virtual teaching systems, and helping to improve the digitalization and intelligence level of vocational skills training.

2.1 Improved image enhancement network

Aiming at the low-light image problem commonly existing in vocational skills virtual teaching scenarios. Such images, due to equipment acquisition conditions or operation environment limitations, often present characteristics such as high noise, low contrast, and blurred texture details, which seriously affect the recognition and localization of target objects. This study designs an improved image enhancement network as the core module of preprocessing. The module architecture is shown in Figure 1.

1.png

Figure 1. Architecture of the improved image enhancement module for vocational skills virtual teaching

This module takes low-light teaching images as input, focuses on the low pixel value region that occupies a high proportion in the image gray histogram, and designs a light enhancement curve for pixel-wise mapping enhancement. By analyzing the typical poor lighting environments in vocational skills teaching scenarios, a nonlinear mapping-based enhancement strategy is proposed, focusing on improving the dynamic range of brightness in low pixel value regions, enhancing the overall brightness while preserving image texture details, and providing clearer visual input for subsequent target detection. Specifically, let the input pixel value be represented by A, the output pixel value by B, and the light enhancement parameter to be learned by β. The mapping formula between input and output pixels is:

$B(A ; \beta)=A+\beta\left(A-A^v\right)$            (1)

To solve the problems of pixel overflow and insufficient dynamic range that may occur during the enhancement process of virtual teaching images, the image pixel values are first normalized to the [0,1] interval. A nonlinear mapping curve is constructed by introducing the adjustable parameter β and the hyperparameter v. Among them, β controls the overall offset of the curve to adjust image brightness, and the high-order design of v can realize a steeper mapping slope in the low pixel value region, so that dark pixels can be mapped to a wider output value range, effectively improving the recognition of details in dark areas. Although high-order curves may sacrifice some monotonicity in high pixel value regions, in the actual enhancement of vocational skills teaching images, this strategy has a suppression effect on local overexposure caused by equipment reflection, and is especially suitable for strong light direct or backlight scenarios commonly seen in operation videos, ensuring that the brightness adjustment of key targets such as tool edges and instrument scales is not overexposed and clearly presented. It allows the image to be adjusted in a wider dynamic range. The expression for the output image A of the s-th enhancement is:

$A_a\left(A_{a-1} ; \beta\right)=A_{s-1}+\beta\left(A_{s-1}-A_{s-1}^v\right)$             (2)

Considering the diversity of target objects and the unevenness of local lighting in vocational skills teaching images, traditional global parameter enhancement is difficult to meet the personalized needs of different regions. This study expands the global enhancement parameter β into a pixel-level parameter matrix X_v, establishing a correspondence between each light enhancement curve and all pixels in the image, realizing pixel-wise quality enhancement and brightness adjustment. By taking the output of the previous enhancement stage as the input of the next stage, an iterative enhancement mechanism is formed, allowing the model to dynamically adjust the enhancement parameters according to the local features of each pixel in the image. It retains key details such as tool surface textures and equipment indicator light colors, while avoiding noise amplification in the background area caused by over-enhancement, significantly improving the overall quality of teaching images in complex scenarios. Specifically, X_v is divided along the channel dimension into X_v~X_S, which are used as light enhancement parameters for each iteration unit. The expression for X_s is:

$A_s\left(A_{a-1} ; X_s\right)=A_{s-1}+X_s\left(A_{s-1}-A_{s-1}^v\right)$               (3)

The image enhancement module adopts an architecture design of "lightweight parameter prediction network + cascaded iteration units", which is suitable for the real-time and computational efficiency requirements of vocational skills virtual teaching systems. Specifically, the low-light image is simultaneously input into the enhancement parameter prediction network and the first iteration unit: the former estimates a curve parameter matrix X_v consistent with the image size through a lightweight convolutional neural network, and divides it into A₁~A_S by channel dimension to provide pixel-wise enhancement parameters for each iteration unit; the latter performs cyclic processing on the input image through S cascaded image enhancement units, with each level unit using the previous output as input and performing light enhancement mapping based on the corresponding parameter matrix X_s. This cascaded iteration mechanism allows the model to gradually adjust image brightness and contrast over a wider dynamic range, especially suitable for virtual teaching scenarios where lighting conditions change frequently during operation. The final output enhanced image not only has normal brightness and clear texture, but also highlights the distinction between foreground targets and the background, providing high-quality feature input for the subsequent YOLOv8 object detection network, effectively improving the accuracy and real-time performance of target detection in complex teaching scenarios.

2.2 Enhancement parameter prediction network

Aiming at the complex characteristics of low-light images in vocational skills virtual teaching scenarios, such as local shadows, uneven illumination caused by equipment reflections, and multi-scale feature differences of target objects like tools and operation parts, the enhancement parameter prediction network achieves accurate prediction of adaptive light enhancement curve parameters through a hybrid dilated convolution and cross-layer connection mechanism. The network architecture is shown in Figure 2. The network takes the original low-light teaching image as input, adopts a 7-layer convolutional architecture and removes downsampling and batch normalization layers to fully preserve the image spatial dimension information and inter-pixel smooth relationships, which is especially suitable for maintaining the continuity of key details such as edges of operation tools and device scales. In the first three layers, parallel hybrid dilated convolution layers with dilation rates of 1, 2, and 3 are introduced to capture local textures and global illumination distribution of the image through receptive fields of different scales. Multi-scale features are fused through channel concatenation, effectively handling problems of large target size differences and complex lighting conditions in teaching scenarios. By halving and then concatenating the output channels of the first three layers, the network maintains model representation ability while reducing about 50% of the parameter amount, significantly improving computational efficiency and meeting the real-time requirements of virtual teaching systems.

The network outputs S×3 parameter matrices with the same size as the input image through S×3 3×3 convolution kernels. Each matrix element represents the light enhancement curve parameter at the corresponding pixel position, realizing pixel-wise and channel-wise dynamic adjustment. The Tanh activation function is used to constrain the parameter values within the range of [-1,1], ensuring the stability and interpretability of the enhancement curves. Through the above architecture, the network can adaptively generate personalized enhancement strategies based on the characteristics of vocational skills teaching images—for example, generating more aggressive brightness enhancement curves in tool operation areas while applying gentle adjustments in background areas to avoid noise amplification. By applying the output parameter matrices to the S cascaded image enhancement units, progressive optimization of low-light images is achieved, which is especially suitable for handling complex lighting scenarios common in virtual teaching. The final output is an enhanced image with uniform brightness, clear texture, and prominent targets, providing high-quality input for subsequent object detection.

2.png

Figure 2. Architecture of the enhancement parameter prediction network

2.3 No-reference loss function

Aiming at the practical issue of the difficulty in obtaining a large amount of manually annotated or paired datasets in vocational skills virtual teaching scenarios, the no-reference loss function system designed in this study focuses on the intrinsic correlation between the image's own features and pre- and post-enhancement, constructing an unsupervised training framework. This allows the enhancement parameter prediction network to achieve zero-reference learning by analyzing pixel spatial relationships and color exposure characteristics from a single image. The constructed loss function is especially suitable for diverse scenarios in vocational skills teaching images. For example, low-light images taken at narrow angles inside equipment, or local shadow images caused by hand occlusion during operation, can be adaptively enhanced in key regions such as tool contours and operation details without relying on external annotations, avoiding the strong dependency of traditional supervised learning on data labeling and significantly reducing the technical deployment threshold of virtual teaching systems.

(1) Spatial consistency loss (M_SPA)

Aiming at the problem that the edge textures and spatial structures of target objects in vocational skills teaching images are prone to being lost due to over-smoothing during enhancement, the spatial consistency loss divides the image into 4×4 pixel blocks and enforces the consistency of differences in adjacent regions between the input and enhanced images. For example, in a mechanical assembly teaching scenario, this loss ensures that the contrast between the serration of a screwdriver and the groove of a screw is not weakened after enhancement, avoiding blurred operation details caused by local over-enhancement, thus preserving clear geometric feature information for the subsequent object detection network. Its core principle is to maintain spatial boundaries between tools and background, operation and non-operation regions by constraining block-level differences, preventing artifacts or detail distortion during enhancement. Assuming the number of pixel blocks in the image is denoted by J, the average pixel value of the u-th pixel block in the low-light image and the enhanced image is denoted by B_u and U_u, and the average pixel values of the four adjacent regions (top, bottom, left, right) are denoted by B_k and U_k, then the expression is:

$M_{S P A}=\frac{1}{J} \sum_{u=1}^J \sum_{k \in \Psi(u)}\left(\left|B_u-B_k\right|-\left|U_u-U_k\right|\right)^2$             (4)

(2) Exposure control loss (M_EXP)

Aiming at the common problem of uneven local exposure in virtual teaching, such as overexposure on equipment surfaces caused by direct strong light on the operation table, or underexposure in corner areas, the exposure control loss measures the difference between the local exposure level of the enhanced image and a preset good exposure level R, dynamically adjusting the brightness compensation of each region. For example, in electrical operation teaching videos, this loss can avoid detail loss in reflective areas of wire insulation layers caused by over-enhancement, while improving the visibility of dark areas inside the electrical box, making key information such as circuit breaker switch status clearly identifiable. Through region-by-region exposure constraints, the enhanced image is ensured to maintain appropriate brightness in key operation areas, avoiding both overexposure and underexposure, providing balanced visual input for object detection. Assuming the number of non-overlapping local regions of size 16×16 pixels is denoted by L, and the average intensity value of a local region in the enhanced image is represented by B, then the expression is:

$M_{E X P}=\frac{1}{L} \sum_{j=1}^L\left|B_j-R\right|$         (5)

(3) Color constancy loss (M_COL)

Considering the critical role of color features in target recognition in vocational skills teaching images—such as color-coded pipeline identifiers and indicator light statuses—the color constancy loss aims to correct color deviations caused by uneven illumination and maintain a reasonable mapping relationship among the RGB channels. For example, in chemical process simulation teaching, this loss ensures the consistency of pipeline medium color under different lighting conditions, avoiding color distortion caused by channel imbalance during enhancement that may affect learners’ judgment. By establishing inter-channel adjustment constraints, the enhanced image colors become closer to real operation scenes, ensuring the accurate transmission of color-sensitive information such as equipment status and tool type, and providing reliable color features for the object detection model. Assuming the average intensity value of channel o in the enhanced image is denoted by K^o, and the pairwise combinations of the RGB color channels are denoted by γ={(E,H),(E, Y),(H,Y)}, then the expression is:

$M_{\text {COL }}=\sum_{\forall(o, w) \in \gamma}\left(K^o-K^w\right)^2$          (6)

(4) Illumination smoothness loss (M_snX)

Aiming at the possible pixel value mutation in enhanced virtual teaching images—such as speckles caused by noise amplification and artifacts in unevenly illuminated regions—the illumination smoothness loss constrains the smooth variation of neighboring pixels in the curve parameter matrix to ensure spatial continuity of brightness adjustment during enhancement. For example, in welding process teaching videos, this loss avoids brightness discontinuities between welding spots and base materials caused by local over-enhancement, maintaining the natural transition texture of metal surfaces and preventing the object detection model from misidentifying texture noise as defects. By enforcing local monotonicity of the parameter matrix, the loss effectively suppresses high-frequency noise possibly occurring during enhancement, improving the overall visual quality of the image and providing a stable feature foundation for subsequent object detection. Assuming the number of iterations is denoted by V, the horizontal/vertical gradient values of pixel values in the corresponding channels are denoted by $\nabla a$ and $\nabla b$, and the three-color channels of the image are denoted by σ= {R,G,B}, then the expression is:

$M_{s v X}=\frac{1}{V} \sum_{v=1}^V \sum_{z \in \sigma}\left(\left|\nabla a X_v^z\right|+\left|\nabla b X_v^z\right|\right)^2$             (7)

The total loss function M_TO is formed by weighted fusion of the above four losses, constructing a multi-dimensional evaluation system for enhancement effectiveness. Assuming the weight parameters are denoted by Q_TO and Q_snX, the expression is:

$M_{T O}=M_{S P A}+M_{\exp }+Q_{C O L} M_{C O L}+Q_{s n X} M_{s n X}$                 (8)

This loss function does not rely on paired data or manual annotations, and can achieve end-to-end unsupervised training based only on image self-features, making it especially suitable for a large amount of unlabeled practical training video data in vocational skills training. During backpropagation, each loss term collaboratively guides network parameter updates: spatial consistency loss preserves operation details, exposure control loss balances local brightness, color constancy loss corrects color deviation, and illumination smoothness loss suppresses noise. Ultimately, the enhancement parameter prediction network can adaptively generate light enhancement curve parameters that meet the needs of vocational skills teaching, providing enhanced schemes for low-light teaching images with appropriate brightness, complete details, and accurate colors, fundamentally solving the data bottleneck problem of traditional supervised learning and enhancing the robustness and applicability of virtual teaching systems.

2.4 YOLOv8 object detection algorithm

Aiming at the characteristics of vocational skills virtual teaching scenarios where target objects have diverse shapes, large scale variations, and high real-time detection requirements, this study adopts a lightweight improved YOLOv8n as the core framework for object detection. While retaining its efficient detection speed, the model structure and parameters are optimized according to the characteristics of teaching images. First, the backbone network continues to use the CSPDarkNet53 basic architecture, replacing the traditional C3 structure with the C2f module, enhancing cross-layer feature concatenation and gradient flow transmission capabilities, effectively capturing the detail features of target objects in operation videos, such as tool surface textures and device dial scales. The fast spatial pyramid pooling layer (SPPF) fuses multi-scale spatial features through multi-layer max pooling with 5×5 kernels and channel concatenation, which is especially suitable for handling high-resolution targets in close-up operation areas and low-resolution equipment in distant environments in teaching images, ensuring accurate detection of targets at different distances.

In response to the problems of blurred object edges and color distortion caused by complex lighting conditions in virtual teaching images, the decoupled head structure of YOLOv8 plays a key role: after discarding the objectness branch, the classification and regression branches are optimized independently, avoiding target confidence misjudgments caused by uneven illumination. The regression branch introduces Distribution Focal Loss (DFL) integral representation, enhancing the positioning accuracy for irregular targets, which is particularly suitable for detecting non-standard geometric shaped objects in scenarios such as mechanical repair and electrical operation. In addition, the FPN+PAN structure in the Neck part fuses multi-level semantic information from the backbone network while strengthening low-level localization features, compensating for possible local detail blurring caused by image enhancement. By adjusting the data augmentation strategy, the model achieves real-time and accurate detection of targets such as tool operation areas and key device interfaces on enhanced low-light images, providing reliable visual information support for functions such as automatic operation evaluation and error action recognition in virtual teaching systems, effectively improving the intelligent interaction level of vocational skills training. Figure 3 shows the object detection model architecture for vocational skills virtual teaching. Figure 4 shows the adopted YOLOv8 network architecture.

3.png

Figure 3. Object detection model architecture for vocational skills virtual teaching

4.png

Figure 4. Adopted YOLOv8 network architecture

2.5 Regression loss function

In vocational skills virtual teaching scenarios, object detection faces challenges such as blurred tool edges under complex lighting and inaccurate bounding box localization caused by occlusion in operation areas. The classification loss commonly used in YOLOv8 is VFL Loss. Assuming the predicted probability of a sample is denoted by o∈[0,1], the expression is:

$\begin{aligned} & \operatorname{VFS}(o, w)= \left\{\begin{array}{l}-w(w \log (o)+(1-w) \log (1-o)), w>0 \\ -\beta o^{\varepsilon} \log (1-o), w=0\end{array}\right.\end{aligned}$          (9)

To measure the overlap between the predicted box and the ground truth box, object detection tasks often introduce IoU. Assuming the intersection and union area of the predicted box and ground truth box are denoted by X∩Y and X∪Y respectively, the expression is:

$I o U=\frac{|X \cap Y|}{|X \cup Y|}$            (10)

The loss function expression corresponding to IoU can be written as:

$M_{I o U}=1-I o U$           (11)

Specifically, assuming the center coordinates of the predicted box and ground truth box are denoted by y and y^hs respectively, the Euclidean distance between the two coordinates is denoted by ϑ²(·), and the diagonal length of the minimum enclosing rectangle of the predicted box and ground truth box is denoted by z, the original CIoU Loss in YOLOv8 can be expressed as:

$M_{Z I o U}=1-I o U+\frac{\vartheta^2\left(y, y^{h s}\right)}{z^2}+\beta n$           (12)

The variables x and n used to measure the aspect ratio difference between the predicted box and the ground truth box can be calculated as follows:

$n=\frac{4}{\tau^2}\left(\operatorname{ARCTAN} \frac{q^{h s}}{g^{h s}}-\operatorname{ARCTAN} \frac{q}{g}\right), \beta=\frac{n}{(1-U p I)+n}$

Although the original CIoULoss of YOLOv8 introduces distance and aspect ratio penalty terms on the basis of IoU, it has a problem of penalty invalidation under geometrical similarity. For example, in electrical operation teaching, screwdrivers of different specifications may fall into local optima during regression due to similar aspect ratios. To address this problem, the improvement strategy is to use EIoULoss as the basis and replace the aspect ratio penalty term in CIoU with M_ASP, which supervises the decoupled width-height difference, directly minimizing the absolute error of width and height between the predicted box and the ground truth box. This improvement is particularly critical in mechanical assembly teaching scenarios, effectively improving the localization accuracy of similar-shaped targets of different sizes, such as nuts and bolts, avoiding regression direction conflicts caused by aspect ratio constraints.

The improved EIoULoss forms multi-dimensional constraints on the predicted box by introducing distance loss M_DIS and side length loss M_ASP: M_DIS normalizes the center point distance between the predicted box and the ground truth box, solving the gradient disappearance problem of traditional IoU in non-overlapping regions, especially suitable for scenarios where targets are partially occluded in virtual teaching, such as only partial contours of tools are visible during hand operation; Lasp separately calculates the square differences of width and height, enabling the network to independently optimize the size parameters of bounding boxes. In welding process teaching images, this decoupled design can more accurately locate weld point areas, avoiding fitting deviation of elliptical predicted boxes to circular weld points caused by aspect ratio penalties. By integrating M_DIS and M_ASP into IoU calculation, EIoULoss realizes comprehensive constraints on bounding box regression, significantly improving the stability of target localization in complex teaching scenarios. Assuming the width of the predicted box and the ground truth box are denoted by q and q^hs respectively, and the height by g and g^hs respectively, and the width and height of their minimum enclosing rectangle are denoted by q^z and g^z respectively, EIoULoss is expressed as:

$\begin{aligned} & M_{\text {RIoU }}=M_{\text {IoU }}+M_{\text {DIS }}+M_{A S P}= 1-I o U+\frac{\vartheta^2\left(y, y^{h s}\right)}{\left(q^z\right)^2+\left(g^z\right)^2}+\frac{\vartheta^2\left(q, q^{h s}\right)}{\left(q^z\right)^2}+\frac{\vartheta^2\left(g, g^{h s}\right)}{\left(g^z\right)^2}\end{aligned}$           (13)

To address the imbalance problem in virtual teaching images where there are a large number of low-error high-quality samples and a small number of high-error low-quality samples—for example, most frames in an operation video have clear tool positions, but only a few frames have localization difficulties due to motion blur—the improvement strategy incorporates the idea of Focal Loss into EIoULoss. By introducing a modulation factor, the gradient contribution of high-quality samples is reduced, while the training weight of low-quality samples is increased, enabling the model to focus more on learning difficult samples during regression. In automotive repair teaching scenarios, this mechanism can effectively improve the detection accuracy of complex pipelines and hidden components inside the engine compartment, reducing false detections caused by local occlusion or uneven lighting. Assuming the hyperparameter used to control the curvature of the loss curve is denoted by ε, the Focal EIoU Loss expression is:

$M_{F-E}=I o U^{\varepsilon} M_{R I o U}$                (14)

This study, targeting the issues of image quality degradation and object detection difficulty in vocational skill training virtual teaching systems, constructed a “image enhancement-object detection” collaborative optimization technical framework. At the image enhancement level, an improved lightweight convolutional network and a non-reference loss function system were used to achieve unsupervised adaptive enhancement of low-light, high-noise teaching images. The enhancement parameter prediction network can automatically generate pixel-level light enhancement curve parameters, effectively improving tool texture clarity and color constancy. At the object detection level, the optimized FocalEIoULoss and decoupled head structure based on YOLOv8 solved the problem of insufficient detection accuracy caused by varied object shapes and light sensitivity in teaching scenarios, significantly improving small object detection mAP@0.5 and meeting the interaction needs of virtual teaching systems.

The theoretical value of the research lies in: for the first time, combining unsupervised enhancement and lightweight detection models, proposing a visual perception paradigm applicable to vocational skills scenarios, providing a new idea for cross-task collaboration of image enhancement and object detection in the education field; the application value lies in: through technical implementation, the virtual teaching system can realize real-time enhancement and tool localization in operation videos, supporting automatic evaluation and intelligent interaction, and promoting the transformation of vocational skill training from “experience-driven” to “data-driven”. However, there are still limitations in the research: First, the enhancement network may still suffer from detail loss in extreme backlight scenes; second, the generalization ability of object detection for rare operational tools needs improvement; third, lightweight deployment of the model on edge computing devices still needs optimization. Future research will be expanded in three aspects: (1) introducing cross-modal fusion to improve robustness under extreme lighting; (2) building a dynamic meta-learning mechanism to enhance rapid adaptability to new category targets; (3) exploring neural architecture search to achieve intelligent matching of model parameters and computing resources, promoting large-scale application of the technology in mobile virtual teaching scenarios.