Deep learning-based novel aluminum furniture design style recognition and key technology research

3.2.1 CNN

CNN is a deep learning model specifically designed for processing data with a grid-like structure (especially images). The structure of CNN mainly consists of the following key layers: convolutional layer, pooling layer, activation function layer, fully connected layer, batch normalization layer, and dropout layer; the entire CNN structure starts from the input layer, through a series of combinations of convolutional, pooling, and activation layers, gradually extracting increasingly more abstract features until the last layer is usually a softmax layer for multi-class classification or a linear layer for regression tasks. The whole CNN structure starts from the input layer and goes through a series of convolutional, pooling, and activation layers to extract increasing abstract features, until the last layer is usually a softmax layer for multi-class classification or a linear layer for regression tasks. The design of this hierarchical structure allows CNNs to automatically learn high-level, abstract feature representations from the raw input data, which has led to great success in several computer vision fields such as image classification, object detection, and semantic segmentation.

This study chooses to apply deep CNN as the core algorithmic tool in this project, aiming to extract high-level abstract representations that can reflect the key design styles and subtle features of aluminum furniture from the rich aluminum furniture image data. Given that our dataset contains aluminum furniture with different design styles and their unique visual elements, we do not directly apply an off-the-shelf CNN model architecture, but carefully tailor a highly targeted and adaptable CNN model based on the characteristics of these data.

3.2.2 Framework design

In this study, CNN was chosen as the core algorithm because it can effectively and automatically extract multi-level abstract features from image data, especially for image data with a grid structure. The aluminum furniture design style recognition task involves identifying different design styles and unique visual elements from images, and CNN can capture local features in images and gradually extract higher-level semantic information through layer-by-layer convolution operations. Therefore, CNN provides an efficient and automated solution for this task, which can automatically learn rich features from the original image without the need for manual feature extraction. Furthermore, the combination of a convolution operation with ReLU activation and pooling effectively forms a multi-level and multi-scale feature representation in the model, which can process aluminum furniture images of different design styles and improve the accuracy and robustness of the model in style recognition. Therefore, CNN was chosen as the main method of this study based on its excellent performance in computer vision tasks, especially in dealing with complex image classification tasks.

CNN Model Architecture Design: Our model commences with a standardized input layer accepting images of dimensions ( 224 × 224 × 3 ) , preprocessed by subtracting the mean and dividing by the standard deviation to ensure input consistency. Sequentially, the model extracts features through a series of meticulously configured convolutional layers, where each layer applies a convolution operation with a kernel W l on the preceding layer I l − 1 , adds bias b l , and introduces non-linearity via the ReLU activation function, represented by C l = f ( W l ⁎ I l − 1 + b l ) . Max-pooling layers are introduced to downsample and retain critical features, defined as selecting the maximum value within each pooling region. Fully connected layers integrate these features, forming a vector Y = W F C ⋅ X + b F C , ultimately transforming outputs into class probability distributions through the Softmax function with p i = e z i ∑ j = 1 K e z j , where z i denotes the logit for class i, and K is the total number of classes.

This study received an image of the original aluminum furniture, assuming that the size of the image is H × W × C, where H is the height, W is the width, and C is the number of channels, usually 3 (RGB).

Convolutional layer: Several convolutional kernels are used to perform convolutional operations on the input image to extract local features of the image and output the feature map. Assuming that the size of the convolution kernel is K × K, the number of kernels is F, the step size is S, and the padding is P, then the size of the output feature map is H − K + 2 P S + 1 W − K + 2 P S + 1 F . The formula of the convolution layer is O i , j , f = ∑ m = 0 K − 1 ∑ n = 0 K − 1 ∑ c = 0 C − 1 I × S × i + m , S × j + n , c × W m , n , c , f + b f , where O i , j , f is the value of the ith row, jth column, and fth channel of the output feature map, I x , y , c is the value of the xth row, yth column, and cth channel of the input image, and W m , n , c , f is the weight of the fth convolution kernel in the mth row, nth column, and cth channel, and b f is the bias term of the fth convolution kernel.

Activation layer: The output feature map of the convolutional layer is non-linearly transformed to increase the expressive ability of the model and output the activated feature map. Commonly used activation functions are ReLu, Sigmoid, Tanh, and so on. The formula for the activation layer is A i , j , f = g ( O i , j , f ) , where A i , j , f is the value of the ith row, jth column, and fth channel of the activated feature map, O i , j , f is the value of the ith row, jth column, and fth channel of the output feature map of the convolutional layer, and g is the activation function. O i , j , f is the output feature map of the convolutional layer, row i, column j, the value of channel f, and g is the activation function.

Pooling layer: The activated feature map is downsampled, the dimension of the feature map is reduced, the important features are kept, and the pooled feature map. Commonly used pooling methods are maximum pooling, average pooling, and so on. Assuming that the size of the pooling window is R × R and the step size is T, the size of the output pooled feature map is T H − R + 1 × T W − R + 1 × F . The formula of the pooling layer is P i , j , f = h ( A × T i : T i + R , T j : T j + R , f ) , where P i , j , f is the value of the ith row, jth column, and fth channel of the pooled feature map, A x : y , z : w , f is the sub-matrix of the activated feature map from the xth to yth rows, zth to wth columns, and fth channel, and h is the pooling function.

The activation layer plays a critical role in introducing non-linearity to the network by applying an activation function to the feature map generated by the convolutional layer. Without this non-linearity, the network would behave like a linear model, limiting its ability to learn complex patterns. Common activation functions include ReLU, Sigmoid, and Tanh. ReLU is particularly favored because it is computationally efficient and helps prevent the vanishing gradient problem. It works by setting negative values in the feature map to zero, leaving only positive values. Sigmoid and Tanh are used when outputs need to be constrained to specific ranges, such as probabilities or symmetric values. By applying these functions element-wise to the feature map, the model gains the ability to capture more sophisticated patterns and representations, which are passed to the next layers in the network for further analysis.

Fully Connected Layers: The pooled feature map is flattened into one-dimensional vectors, and then several fully connected layers are used to further transform the features and output the final classification result. Assuming that the dimension of the flattened feature vector is D, and the number of categories to be categorized is C, then the dimension of the output classification result is C. The formula for the fully connected layer is: y = W x + b the specific structure of CNN is shown in Figure 3.

Figure 3

Structure of the CNN model.

After completing the deep CNN extraction of aluminum furniture image features, in order to obtain a more complete and fine-grained feature representation, we use FPN for cross-layer feature fusion. The design principle of FPN is to make full use of the feature maps of different layers in the CNN model, and to organically combine the shallow features’ rich detail information of shallow features and the advanced semantic information of deeper features to generate a multi-scale feature pyramid.

To ensure precise recognition of aluminum furniture design styles, we propose a meticulous CNN-based solution, meticulously covering architectural design, training strategies, and performance evaluation criteria. This approach is dedicated to deeply mining design features from images and implementing efficient classification.

3.2.3 Training evaluation metrics

Training Procedure: Training initiates are modeled with preprocessing of the original image dataset, incorporating normalization to mitigate the impact of varying lighting and color. The loss function employs cross-entropy, effectively gauging the discrepancy between model predictions and true labels, expressed as L = − ∑ c = 1 K y c log ( p c ) . Optimization leverages the Adam algorithm, an adaptive learning rate method that dynamically adjusts the learning rate to expedite convergence and enhance model performance, following the parameter update rule W t + 1 = W t − η ⋅ m ˆ t / ( v ˆ t + ϵ ) . Training incorporates a mini-batch strategy (with batch size B = 32 or 64 to strike a balance between computational efficiency and learning stability) and implements early stopping to prevent overfitting, thereby ensuring model performance on the validation set.

Evaluation Metrics: Model performance is assessed based on key indicators, including accuracy Acc = TP + TN TP + TN + FP + FN , precision Prec = TP TP + FP , Recall = TP TP + FN , and F1 score F 1 = 2 ⋅ Prec ⋅ Recall Prec + Recall , where TP, TN, FP, and FN stand for true positives, true negatives, false positives, and false negatives, respectively. These metrics collectively form a comprehensive evaluation system, ensuring that our model accurately identifies various design styles and exhibits high reliability and practicality in real-world applications.

In summary, through a well-designed CNN architecture, meticulous training procedures, and rigorous evaluation criteria, our model not only delves into the nuances of aluminum furniture design but also effectively distinguishes different styles, thus providing the furniture industry with powerful design innovation, market analysis, and renewable energy technology to enhance the user experience.

This study passes the high-level features to the low-level by up-sampling and fusion to get a feature map with high resolution and strong semantics. We can use nearest-neighbor interpolation to up-sample the high-level feature map so that it has the same spatial dimensions as the low-level feature map, and then use element-wise addition to fuse the two feature maps to obtain the top-down path of the feature maps, denoted as {M2, M3, M4, M5}, which correspond to the downsampling multiples of {4, 8, 16, 32} for the input images, respectively. The formula for the top-down path is M l = U ( M l + 1 ) + C l , l ∈ { 2 , 3 , 4 } . The following is an example of a top-down path: M 5 = C 5 , where M l is the lth layer feature map for top-down paths, C l is the lth layer feature map for bottom-up paths, U is the upsampling operation, and + is the element-by-element summing operation.

Horizontal linking means adding some convolutional layers between different layers to enhance the expressiveness and smoothness of the features. This study uses 1 × 1 convolutional layer to downscale the feature map of the bottom-up path so that it has the same number of channels as that of the top-down path, and then use 3 × 3 convolutional layer to smoothen the feature map of the top-down path to obtain the final feature pyramid, which is denoted a s { P 2 , P 3 , P 4 , P 5 } and corresponds to the downsampling multiples of the input image as { 4 , 8 , 16 , 32 } , respectively. The formulas of lateral join are C l ′ = Conv 1 × 1 ( C l ) , l ∈ { 2 , 3 , 4 , 5 } , P l = Conv 3 × 3 ( M l ) , l ∈ { 2 , 3 , 4 , 5 } , where C l ′ is the l-layer feature map of the downsized path and Pl is the l-layer of the final feature pyramid. Pl is the l-layer feature map of the final feature pyramid, and Conv k × k is a k × k convolutional layer.

Finally, this study defines a detection head for object detection and categorization of the feature maps at each level. The detection head is a small CNN that outputs category scores and bounding box regression results for each candidate region. This study directly predicts the coordinates of the bounding box by using the anchorless box.

The centroid domain-based method adds a centrality branch to the feature vector at each location for predicting the centrality score of that location and a regression branch for predicting the distances from that location to the four boundaries of the object: left, top, right, and bottom. The formulas for the centrality domain-based approach are as follows: y = c , if x ∈ C 0 , otherwise . l = x − x min , r = x max − x , l = x − x min , r = x max − x , t = y − y min , b = y max − y , t = y − y min , b = y max − y , where y is the category label of the location, c is the category of the object, C is the centrality domain of the object, x and y are the coordinates of the location, and x min , x max , y min , y max are the coordinates of the boundaries of the objects, and l, r, t, and b are the distances from the location to the boundaries of the object.

In this study, the hyperparameters for the CNN-based aluminum furniture design style recognition model were carefully selected through experimentation and optimization. The initial learning rate was set to 0.001, with a decay factor of 0.96 every 20 epochs to ensure stable convergence. A batch size of 32 was chosen for an optimal balance between training stability and computational efficiency. The model was trained for a maximum of 50 epochs, with early stopping implemented if the validation loss did not improve after 10 consecutive epochs to prevent overfitting. The number of filters in the convolutional layers started with 32 filters in the first layer and progressively increased to 256 filters in the deeper layers, capturing increasingly complex features. A 3 × 3 kernel size was used in most layers, while 5 × 5 kernels were applied in the initial layers to capture broader features. To prevent overfitting, a dropout rate of 0.5 was applied to the fully connected layers. These hyperparameter values were chosen to balance the model's performance, training efficiency, and generalization, resulting in a robust model for accurately recognizing aluminum furniture design styles.

4.2.1 Experimental design

In the model evaluation stage, this study chose an independent test set to conduct a comprehensive and rigorous evaluation of the constructed aluminum furniture design style recognition model in order to verify the model’s generalization ability and classification performance on unknown data. In the evaluation process, we adopted several classical evaluation indexes, specifically including:

1. Accuracy: This is the most intuitive and easy-to-understand assessment criterion and represents the proportion of samples correctly classified by the model out of the total number of samples.

2. Precision: The proportion of all samples predicted by the model to be in a particular category that actually belong to that category. It reflects the reliability of the model’s predictions.

3. Recall: The proportion of all samples that are actually in a given category that are correctly predicted by the model. It reflects the model’s coverage of the target category.

During the data screening process, the research team manually reviewed the design style of each image to ensure that the label of each image was accurate. For some samples with blurred boundaries or design styles that were difficult to clearly classify, we used further manual review and cluster analysis methods to determine their most suitable classification. In addition, during the training process, images that could not be properly input into the network due to excessive size or incomplete images were excluded. In order to ensure the generalization ability of the model, the number of samples in the training data set was balanced to avoid affecting the performance of the model due to too many or too few categories.

First, all image data were manually reviewed to ensure that the design style label of each image was accurate. For some unclear images or samples whose design styles were difficult to clearly classify, the research team used further manual review and re-labeled them in combination with cluster analysis methods to ensure the accuracy of classification.

During the data preprocessing process, some samples with too large an image size or incomplete content were excluded. These images may not be correctly input into the neural network, resulting in increased errors during training, so they were eliminated. In order to avoid the impact of category imbalance on model training and ensure the representativeness of the data set, the number of samples in different design style categories was balanced in the study to avoid bias in model performance caused by too many or too few categories.

In addition, in order to improve the generalization ability of the model, some samples with great variation in image quality or content were excluded, which may cause overfitting and affect the performance of the model on the test set. Through these strict data screening and exclusion measures, this study ensured the quality of data and the reliability of model training.

In addition, in order to better measure the performance of our model, we also compare our self-developed model with several existing well-known deep learning models, including but not limited to the classic CNN model, the recent popular Visual Transformer (vit) model, and potentially other related models such as CMT. Through the comparison experiments, we can understand the competitive advantages and room for improvement of our own model’s competitive advantage and room for improvement in the aluminum furniture design style recognition task, providing a basis for further optimization and iteration. The specific formulas for the evaluation indexes are as shown in equations (1)–(3).

4.2.2 Experimental results

This study uses the pytorch framework to implement our model and the sklearn library to compute the evaluation metrics. The results of our experiments are shown in Tables 3 and 4.

As can be seen from Tables 3 and 4, our model achieves good results in furniture style recognition, with accuracy, precision, recall, and F1 value reaching above 0.93, indicating that the model is able to distinguish different furniture styles effectively. From the confusion matrix, we can see that the model is able to classify correctly on most of the categories, and there is only some confusion on the two categories of modern and Scandinavian, which may be due to the fact that these two styles have some similar features, such as simplicity, brightness, functionality, etc., which makes it difficult for the model to distinguish them.

This study evaluated the four models on the test set and obtained the results shown in Table 5. As can be seen from the table, the CNN novel aluminum furniture design style recognition model proposed in this study outperforms the comparison models in all evaluation metrics, indicating that the model has strong recognition and generalization capabilities. In particular, the model performs most prominently on modern and Japanese styles, reaching 98 and 97% accuracy, respectively, which may be attributed to the fact that these two styles are more easily distinguished by the model because of their more distinctive features and greater differences from other styles. On the contrary, the model performs relatively poorly on European and American styles, with only 86 and 88% accuracy, respectively.

Table 6 compares, through metrics such as “Key Design Style Extraction Accuracy,” “Micro-feature Capture Score,” “Number of Training Dataset Images,” and “Feature Vector Dimension,” the ability of pre-trained CNNs and other models to extract high-level abstract representations from aluminum furniture image data. Pre-trained CNNs demonstrate a high level of accuracy and feature capture capability in this task.

Table 7 quantifies feature fusion effectiveness using Average Precision (mAP) on the COCO dataset and recall rates for objects of different sizes, along with the average latency required for feature fusion. FPN excels not only in mAP but also shows a significant advantage in small object recall, albeit with a relatively higher feature fusion latency.