Application of Image Big Data in the Ecological Agricultural Product Supply Chain

With the rapid development of ecological agriculture, the issue of agricultural product quality detection has become increasingly complex and important. Ecological agricultural products, due to their natural and pollution-free characteristics, usually have high market value and consumer recognition. However, the quality standards for ecological agricultural products often lack unified specifications, and small defects, blemishes, or damages on the surface of agricultural products are difficult to efficiently and accurately assess using traditional manual detection methods. As the supply chain becomes increasingly globalized, agricultural products are easily affected by external factors during transportation, storage, and sales, leading to inconsistencies in product appearance. Therefore, using efficient technical means for quality monitoring, especially through image big data technologies for precise surface analysis of agricultural products, can allow real-time and comprehensive understanding of the quality status of agricultural products and ensure that each link meets quality standards. This image big data-based quality detection method can effectively improve detection efficiency, reduce human errors, and ensure the stability and consistency of product quality, thus enhancing the competitiveness of ecological agricultural products in the market.

Surface disparity estimation, as an important computer vision technology, can precisely measure the details of agricultural product surfaces, capturing surface defects, damages, and other quality issues. Methods based on disparity estimation can effectively simulate and reconstruct the three-dimensional structure of agricultural product surfaces, which is of great significance in determining their surface quality and defects. Compared with traditional image processing methods, deep learning-based surface disparity estimation methods can maintain high detection accuracy even under complex backgrounds and changing lighting conditions. Deep learning technology can automatically learn surface features of agricultural products in different states and types from a large number of agricultural product images, further improving the adaptability and robustness of the detection model.

2.1 Convolutional layer

CNNs have powerful feature extraction capabilities in the field of computer vision, allowing them to automatically learn the complex textures, color changes, and depth information of agricultural product surfaces through multiple convolutional operations, capturing subtle morphological features. Particularly for ecological agricultural products, their surfaces may exhibit natural inhomogeneity, such as the skin texture, roughness, and small-scale concave-convex structures of fruits and vegetables. CNNs can gradually learn surface information at different scales through hierarchical feature extraction mechanisms, and, in combination with disparity estimation algorithms, construct an accurate three-dimensional surface model of agricultural products. Figure 1 shows the schematic diagram of the CNN architecture.

1.png

Figure 1. Schematic diagram of CNN architecture

For the task of surface disparity estimation of agricultural products, the CNN architecture needs to possess key characteristics such as multi-channel feature extraction, hierarchical depth matching, and adaptive feature fusion. Multi-channel feature extraction ensures that the model can capture depth information of agricultural product surfaces from different angles and under varying lighting conditions, avoiding misjudgments caused by lighting changes in traditional 2D image analysis. Hierarchical depth matching, achieved by constructing multi-scale convolutional layers, allows the model to capture large-scale surface curvature while retaining fine local disparity information, enabling accurate identification of minor surface defects of agricultural products. Adaptive features allow the model to dynamically adjust the feature information at different layers, enhancing the model's adaptability to complex surface forms, so that the detection system can be applied to ecological agricultural products of different types and production environments. Specifically, suppose the calculated pixel at position (u, k) in the f-th output feature layer of the m-th convolutional layer is represented by T_f^m, with l and v being the size of the convolution kernel and y_f^m being the bias at the current layer. The convolution operation output can be expressed as:

$T_f^1(u, k, j)=d\left(\sum_{l, v} U^{m-1}(u+l, k+v) \times J_{j f}^m(l, v)+y_f^m\right)$  (1)

2.2 Pooling layer

For the task of surface disparity estimation of agricultural products, the main goal of the pooling layer is to improve the model's sensitivity to important features by removing unnecessary details, reducing noise and redundant information in the image. Since the surface quality differences of agricultural products may be caused by small texture changes or slight surface flaws, the pooling layer helps the model ignore subtle surface variations and focus on more representative regional features, thus avoiding overfitting and improving the model's adaptability to image data in complex environments. Suppose the resampling factor is represented by $\theta_f^m$, the downsampling factor is represented by SU(·), and the bias is represented by y_f^m. The output of the pooling layer can be expressed as:

$o_f^m(u, k, j)=d\left(\theta_f^m S U T_f^m(l, v, j)+y_f^m\right)$   (2)

The pooling layer further enhances the performance of the surface disparity estimation model for agricultural products by using either max pooling or average pooling strategies. Max pooling retains the most significant features in the image by selecting the maximum value within the pooling window, which is particularly effective for capturing prominent defects on the agricultural product surface, such as cracks, stains, or damage. On the other hand, average pooling calculates the average value within the pooling region, integrating more regional information to improve the model's stability and robustness to local features. In surface disparity estimation for agricultural products, this characteristic of the pooling layer is particularly important because quality issues on the surface of agricultural products often involve changes in local areas. Through pooling, the network can better adapt to variations under different scales and surface conditions, thereby enhancing its overall ability to assess the quality of agricultural products.

2.3 Fully connected layer

The surface features extracted progressively through the convolutional and pooling layers are mapped into a higher-dimensional space when they enter the fully connected layer, allowing the network to capture complex feature interactions. For the task of agricultural product surface disparity estimation, the role of the fully connected layer is not just to perform a simple weighted sum of features but to integrate features from different scales and regions through connections with multiple neurons, extracting the correlated features between surface depth information and quality.

To improve the model's stability and generalization performance, the fully connected layer typically employs Dropout strategies and regularization methods to prevent overfitting when handling complex agricultural product surface data. Since surface disparity estimation of agricultural products involves dealing with various environmental lighting and surface texture changes, Dropout helps to effectively prevent the network from relying too heavily on specific neurons’ feature representations during the training process by randomly discarding a certain proportion of hidden layer neurons. This enhances the model's robustness to different agricultural product samples.

2.4 Regression layer

The regression layer, as the final layer, is primarily responsible for predicting the surface depth information of agricultural products based on the features extracted and integrated by the preceding layers, and outputs numerical values that can be used for quality assessment. Since the goal of surface disparity estimation for agricultural products is to acquire precise surface depth information based on image data for identifying surface defects, the regression layer must be capable of handling continuous output problems. By continuously adjusting the model parameters, the regression layer learns the patterns of surface depth distribution of agricultural products, improving the sensitivity to small surface undulations, depressions, or protrusions, thus providing more reliable numerical data for subsequent quality assessment.

In specific agricultural product quality grading or defect classification tasks, the regression layer can also combine activation functions such as Softmax or Sigmoid for discretization, transforming the continuous depth estimation results into discrete category labels. For example, the surface quality of agricultural products can be categorized into multiple levels based on disparity information, such as "Normal," "Minor Defect," "Severe Defect," etc. In this case, a Softmax layer can be used to compute the probabilities of different categories and select the category with the highest probability as the final predicted result. Softmax ensures that the sum of the probabilities of all categories is 1 and provides clear classification boundaries, making the results of agricultural product quality detection more intuitive and easier to interpret. Suppose the number of labels is represented by j, and the Softmax function can be expressed as:

$d_u(b)=\frac{\exp \left(b_u\right)}{\sum_j \exp \left(b_u\right)}$      (3)

Given the input a_u and parameter q, the probability of output label k(b_u=k) can be calculated as follows:

$o\left(b_u=k \mid a_u ; q\right)=\frac{\exp \left(q_k \bullet a_u\right)}{\sum_{v=1}^j \exp \left(q_k \bullet a_u\right)}$   （4）

2.5 Gradient descent and backpropagation

Since the quality detection of agricultural products involves complex image features such as texture changes, cracks, stains, etc., optimizing the loss function using gradient descent can effectively minimize the difference between the predicted values and the actual values. In this process, the loss function is used to measure the deviation between the model's predicted value and the actual value, i.e., the disparity between the depth estimation result and the actual surface depth. The goal of the network is to reduce this gap through gradient descent. Specifically, gradient descent updates the trainable parameters of the network in the opposite direction of the loss function gradient, enabling the model to gradually adjust its weights and biases to more accurately predict the surface depth information of agricultural products. The constructed M₁ loss function is represented as:

$M_1(\hat{b}, b)=\sum_{u=0}^l\left|b^{(u)}-\hat{b}^{(u)}\right|$   （5）

Backpropagation plays a crucial role in the gradient descent process. It uses the chain rule to compute the gradient of each neuron layer by layer and propagates the loss backward from the output layer to the input layer. This process helps compute the gradient of each layer's parameters so that the optimizer can adjust the weights and biases in the network, reducing the output error. In the task of surface disparity estimation for agricultural products, backpropagation efficiently updates the parameters of various network layers such as convolutional layers, pooling layers, and fully connected layers, enabling the model to learn from errors and improve with each training iteration, gradually enhancing sensitivity to small surface defects of agricultural products. The total discrepancy calculated by backpropagation can be expressed as:

$E_{T 0}=\sum \frac{1}{2}(\text { Target value }- \text { Calculated output value})^2$  （6）

End-to-end deep learning stereo matching models are typically optimized directly from raw images to disparity map outputs through end-to-end learning. This type of model has strong global feature learning capabilities and can automatically extract and optimize relevant features from the images. However, when faced with surface details of agricultural products, especially in situations with complex lighting changes, surface textures, and small-scale defects, challenges may arise. This is because end-to-end models tend to optimize performance globally, potentially neglecting certain details, which can lead to insufficient accuracy when processing fine surface defects. On the other hand, pyramid networks provide stronger adaptability in multi-scale feature processing. Particularly when dealing with the diverse defects of agricultural product surfaces, the pyramid structure can refine the localization and estimation of small surface changes through disparity maps at different scales, effectively improving the robustness and accuracy of agricultural product surface quality detection. The loss function defined for the model is as follows:

$M(f, \hat{f})=\frac{1}{V} \sum S O_{M_1}\left(f_u-\hat{f_u}\right)$  (12)

In the disparity computation process, the pyramid stereo matching network uses 3D CNNs to regularize the 4D cost volume, improving the stability of disparity matching. Unlike fixed 3D convolution layers and hourglass structures, the pyramid stereo matching network can adjust features on the cost volume at different scales, making the model more flexible in handling complex backgrounds, lighting changes, and local occlusions. For example, when detecting diseases on the surface of ecological agricultural products, some subtle lesions may be difficult to identify on a disparity map at a single scale. However, the pyramid model can leverage multi-scale information, making the depth information of the diseased areas clearer. During the disparity regression phase, the pyramid network typically employs the softargmin method, calculating the probability distribution of different disparity values through the softmax operation and performing a weighted sum to obtain a smoother and more accurate disparity map. The disparity estimation results can be expressed as:

$\hat{f}=\sum_{f=0}^{F_{M A X}} f \times \delta\left(-z_f\right)$   (13)

This paper addresses the technical bottlenecks in agricultural product surface disparity estimation by proposing a new image analysis method based on deep learning, aimed at improving the detection accuracy and robustness of agricultural product surface defects. The research mainly includes three aspects: First, a CNN for agricultural product surface disparity estimation is proposed, which efficiently extracts detailed features of the agricultural product surface, providing strong support for subsequent defect detection. Second, an end-to-end deep learning stereo matching model is designed, which further enhances the detection capability of agricultural product surface defects and optimizes the accuracy of disparity estimation in traditional methods. Finally, a pyramid stereo matching network model is proposed, which, through the fusion of multi-scale features, enhances the model's adaptability and robustness at different resolutions, effectively solving the problem of feature extraction in low-resolution images. Experimental results show that the proposed method demonstrates significant performance improvements on both the training and validation sets, especially in the accuracy of agricultural product surface disparity estimation, reaching near-perfect levels, validating the method's effectiveness and feasibility.

Overall, the research in this paper has important application value, particularly in the agricultural field for agricultural product quality inspection and automated sorting. Through high-precision surface defect detection and quality state classification, this method is expected to play a key role in improving agricultural product production efficiency, reducing human intervention, and enhancing the accuracy of quality inspection. However, there are certain limitations in this study. First, although the model performs well on the experimental dataset, its generalization ability still needs further validation, especially in diversified and complex real agricultural environments. Second, the computational resources required for model training are relatively large, which may pose some cost pressure for practical applications. Future research directions can focus on the following aspects: First, exploring the model's adaptability and robustness on more diverse datasets to enhance its generalization ability in real production environments. Second, optimizing the model's computational efficiency and reducing reliance on hardware resources to enable application on lower-cost devices. Finally, multimodal fusion with other types of sensor data can be considered to further improve defect detection and quality assessment accuracy and reliability.