In this paper, 4,300 site photos were collected at different construction sites and stages, and the samples were expanded to 8,600 by means of horizontal mirror image data enhancement. A data set containing 24,650 safety helmets of different scales, different light intensities and different shielding conditions was made. The ratio of training set and verification set during the training process was 8:2. Based on the samples in the pilot electric room scenario. The number of markings for pedestrians, work clothes, work caps, gloves, hand-held operations, and poles is shown in Table 1. The data preprocessing steps are as follows.

1) Resampling the training set data.

At present, although the CNN algorithm has achieved a high recognition rate, the computational load is relatively large and does not meet the real-time requirements (Ramcharan et al., 2017). In practical application scenarios, it is not only necessary to take into account the accuracy of the identification of electric power workers’ safety work, but also to consider the real-time nature of the identification, to inform the electric power workers in time and prevent the occurrence of safety accidents (Ker et al., 2017; Mezgec et al., 2019). Therefore, to improve the real-time performance of the CNN algorithm, we need to prune the original network model and propose a lightweight network model. Under the condition of ensuring the same accuracy, the recognition speed of the network model is accelerated, and the real-time performance of the network model is improved. This paper improves the CNN algorithm in the following aspects.

1) Replace the convolution kernels of all convolutional layers with 3*3 convolution kernels. Two 3*3 convolutional layers are equivalent to a 5*5 convolutional layer, and three 3*3 convolutional layers are equivalent to a 7*7 convolutional layer, in the case of the same field of view, the network level is deepened, the non-linear transformation is added, the feature learning ability of the network is stronger, and the network capacity is larger. Compared with the large convolution kernels of 5*5 and 7*7, the number of parameters of the small convolution kernel of 3*3 is significantly reduced.

To protect the feature distribution learned by the network, the network changes are reconstructed and learnable parameters γ and β are introduced to inverse normalization, so y k = γ k x ^ k + β k . The batch normalization algorithm can prevent gradient disappearance or gradient explosion to some extent.

4) Use the ReLU activation function with better performance instead of the Sigmoid activation function, it can be expressed as:

When the Sigmoid activation function is used, there are the following three obvious disadvantages: 1) The network input is too large or small, the neuron gradient will tend to zero, and the neuron gradient will disappear during backpropagation, which will cause the neural network to fail to train; 2) The output means of sigmoid activation function is non-zero, and the non-zero mean signal output by the neurons in the previous layer will be used as the input signal of neurons in the next layer. When the input data is positive, the gradient will always be updated in the positive direction; 3) The calculation of the sigmoid activation function is more complicated, which will increase the network training time for large-scale deep networks (Aggarwal, 2019). The formula of the ReLU activation function is: σ x = max 0 , x (3)

Compared with the Sigmoid activation function, the ReLU activation function performs better and helps in the propagation of gradients. The ReLU activation function has a relatively small amount of calculation, and only needs to do one arithmetic operation. The ReLU activation function is always 1 for the part greater than 0, and the gradient will not be saturated (the gradient will not be too small). During the backpropagation process, the gradient can be better propagated to the previous network, and the network will converge faster. The improved structure parameters of the CNN model are shown in Table 2.

About the improved CNN model, the input layer is a 32*32 work picture of electric power workers. There are four convolutional layers, the activation functions are all ReLU, the stride is 1, the padding equal to the same, and zeros are filled around the input picture. The size of the convolution kernels of the first and seventh layers of convolutional layers are both 3*3, and the number of convolutional kernels is 64 and 256 respectively. The third and fifth layers of convolutional layers are Inception modules, which consist of four parts.

The number of neurons in the last convolutional layer is set to 128, 256 and 512 respectively, and the experimental results under different numbers of convolutional neurons are compared and analyzed, as shown in Table 3. According to the results, when the number of neurons in the last convolutional layer is 256, the recognition rate of the model is the highest, so the number of neurons in the last convolutional layer is selected as 256.

To select the better optimizer, the optimizers of the stochastic gradient descent algorithm (SGD) (Sharma, 2018), Momentum algorithm (Li et al., 2021), Adagrad algorithm (Traoré and Pauwels, 2021), RMS prop algorithm (Xu et al., 2021), and Adam algorithm (Jais et al., 2019) are compared with the gradient descent algorithm. Table 4 shows the performance of the network model in different optimization algorithms.

In Table 4, when the network model adopts the Adam optimizer, the training time of the network model is the shortest and the recognition rate is the highest, so the Adam optimizer is used to identify the safety wear of electric power workers.

The loss function in the algorithm of safe wear recognition is designed as follows: L s , y = 1 n ∑ t = 1 n s t − y t 2 (4)

In 4), y is the label value; s is the predicted value of the network forward propagation.

The back-propagation process is as follows: ∂ L ∂ s = 2 n s 1 − y 1 , … , s n − y n (5)

To evaluate the recognition effect, the average accuracy index is used to measure the matching accuracy of the detection frame to the target object, it can be expressed as follows (Chen et al., 2022b): A P i = ∑ N   TruePositives   i N   Totalobjects   i N   TotalImages   i (6) where, N   TruePositives   i is the actual quantity of pictures at ith category; N   Totalobjects   i is the quantity of all objects at the data set category i; N   TotalImages   i is the number of images containing objects in the category i.

The average precision metric is: m A P = ∑ i N A P i / N (7)

During the network training process, the early-stopping mechanism is used to prevent the network from overfitting, and the parameter is set to 15. Training is stopped when the loss on the training set drops while the loss on the validation set remains the same for 15 consecutive epochs. To avoid the network from overfitting, the weight of the network model is only taken as a small value to limit the complexity of the network model and make the weight distribution more regular. This is weight regularization, adding the cost associated with larger weights to the network loss function to make the absolute value of the weight coefficients small enough. Based on the L2 regularization method (Chen and Zhao, 2021; Wang et al., 2021; Liu et al., 2023), the formula is as follows: J = J 0 + λ 2 ‖ w ‖ 2 2 (8)

The training parameters are set as shown in Table 5.

The training time of the improved CNN network model is 1531s. The accuracy curve of the network training curve is shown in Figure 2 and the loss change curve of the network training curve is shown in Figure 3.

In Figures 2, 3, we can see that the accuracy in the early stage of training gradually increased, and the loss of the training set and the validation set also gradually decreased. There was a slight oscillation in the middle, and it gradually became stable with iterations. The network accuracy eventually tends to 100%, and the network loss eventually tends to 0. When the epoch is equal to 14, the loss of the training set is still decreasing, the loss of the validation set tends to remain unchanged, and when the loss of the validation set does not change for 10 consecutive times, the early-stopping mechanism will end the training of the network in advance, so the final training is taken. The network model was obtained 25 times. The safety wear recognition results of the electric staff working are shown in Table 6.

To verify the superiority of the improved CNN algorithm, the network model proposed in this paper is compared with the other network model under the same parameters. The performance comparison result is shown in Table 7.

In Table 7, the recognition method used in this paper has the highest recognition accuracy, reaching 96%. The highest recognition rate of other algorithms is 92.5%. Although these recognition methods solve the problem of image recognition to some extent, when the image features are too mixed, these algorithms cannot well complete the mapping from feature extraction to state recognition, which is not conducive to model recognition. Compared with other traditional algorithms, the improved deep learning algorithm has a higher recognition rate and significantly improves the recognition rate of safe wearing of electric power workers. Compared with other algorithms, the proposed method has faster running speed and can meet the requirements of real-time computing. The proposed network model has strong feature expression ability. The advantages of good generalization performance and strong robustness.