Commodity identification device and method based on improved convolutional neural network

Abstract

The invention discloses a commodity recognition device and method, and particularly relates to a commodity recognition device and method based on an improved convolutional neural network, and the device comprises a data collection module which comprises a bottom plate, the bottom plate is provided with a first track, the first track is provided with a first sliding block in a relative sliding manner, and the first sliding block is provided with a second sliding block; a second rail is fixed to the first sliding block, and a second sliding block is arranged on the second rail in a relative sliding mode. An improved convolutional neural network model is adopted to carry out commodity identification, the model can extract feature information in a picture more accurately through a feature extraction module, a feature transformation module and a classification module, and efficient classification and identification are carried out. The model structure optimizes the recognition algorithm and improves the recognition accuracy and efficiency.

Description

Technical Field

[0001] The present invention relates to the field of new generation information technology, and in particular to a commodity recognition device and method based on an improved convolutional neural network. Background Art

[0003] In supermarkets and e-commerce, product recognition is a crucial technology. Traditional product recognition methods often rely on manual operations, such as barcode scanning or manual input of product information. However, these methods have many limitations, such as the easy damage of barcodes, inefficient manual input and easy errors. With the advancement of science and technology and the rapid development of deep learning technology, product recognition technology based on computer vision has gradually emerged and become an effective means to solve these problems. As an important model in deep learning, convolutional neural network (CNN) has shown excellent performance in the field of image recognition. Through multi-layer convolution and pooling operations, it can automatically extract feature information from images and abstract layer by layer, and finally realize accurate recognition of image content. In recent years, CNN-based product recognition technology has received widespread attention and research, and has made significant progress. However, the existing CNN-based product recognition technology still has some shortcomings. For example, the accuracy of product recognition in complex scenes needs to be improved. Therefore, how to further optimize the CNN model and improve the accuracy and speed of product recognition has become a hot spot and difficulty in current research. Summary of the invention

[0004] In order to solve the above-mentioned problems in the prior art, the present invention provides a commodity identification device and method based on an improved convolutional neural network to solve the current technical problems.

[0005] The technical solution adopted by the present invention to solve its technical problem is: The present invention provides a commodity identification device and method based on an improved convolutional neural network, including: a data acquisition module, the data acquisition module including a base plate, a first track is arranged on the base plate, a first slider is arranged on the first track for relative sliding, a second track is fixed on the first slider, and a second slider is arranged on the second track for relative sliding.

[0006] Preferably, a baffle is provided at the end of the first track, and a first telescopic rod for driving the first slide block to slide on the first track is provided between the baffle and the first slide block.

[0007] Preferably, a second telescopic rod is provided between the first slider and the second slider to drive the second slider to slide on the second track.

[0008] Preferably, a camera is arranged on the second sliding block, a layer of iron sheet is pasted on the back of the second track, a magnet is adsorbed on the iron sheet, a connecting rod is connected to the magnet, and a fill light is arranged on the connecting rod.

[0009] Preferably, a plurality of springs are provided on the bottom plate, and a storage plate is connected to the springs.

[0010] Preferably, a plurality of proximity switches are provided on the bottom plate, and the proximity switches are provided below the storage plate.

[0011] Preferably, the identification method includes the following: Step 1: Place the product to be identified on the shelf, trigger the proximity switch, and the camera takes a photo; Step 2: Input the photos into the improved convolutional neural network model to classify and identify the products; Step 3: Output the classification results of classification recognition.

[0012] Preferably, in Step 1, the camera can be filmed by triggering one of the proximity switches, and the triggering modes of the proximity switch include a first triggering mode and a second triggering mode. The first triggering mode is triggered by the mass of the product itself when the gravity of the product itself is greater than the critical pressure of the placement plate to overcome the spring pressure and trigger the proximity switch; the second triggering mode is triggered by the user manually pressing the placement plate when the gravity of the product itself is less than the critical pressure of the placement plate to overcome the spring pressure and trigger the proximity switch, and the critical pressure is 10-50N.

[0013] Preferably, in Step 2, improving the convolutional neural network model includes: Feature extraction module: includes several convolutional layers and pooling layers to extract image information and generate feature maps; Feature transformation module: performs feature transformation on the feature map extracted by the feature extraction module; Classification module: Use a multi-layer perceptron to process the data and obtain the final classification result.

[0014] The beneficial effects of the present invention are: Flexibility and adaptability: By designing a structure with sliding tracks and sliders, the device can flexibly adjust the position of the camera to better adapt to the shooting needs of goods of different sizes and shapes. This design improves the adaptability of the device to goods, making the identification process more efficient and accurate.

[0015] High-precision recognition: An improved convolutional neural network model is used for product recognition. Through the three modules of feature extraction, feature transformation and classification, the model can more accurately extract feature information from images and perform efficient classification and recognition. This model structure optimizes the recognition algorithm and improves the accuracy and efficiency of recognition.

[0016] Ease of use and scalability: The entire device has a simple structure, convenient operation, and is easy to maintain and upgrade. At the same time, by replacing or adjusting components such as the camera, the function and application scope of the device can be further expanded to meet the needs of commodity identification in different scenarios.

[0017] High efficiency and energy saving: Through intelligent trigger shooting mechanism and efficient recognition algorithm, the device can complete the identification of goods in a short time, reducing the waste of energy and time. At the same time, in idle state, the device can automatically enter low power consumption mode to further reduce energy consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The above aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which: Figure 1 is a schematic diagram of a commodity identification device and method based on an improved convolutional neural network according to an embodiment of the present invention; Figure 2 is a structural schematic diagram of a commodity identification device and method based on an improved convolutional neural network according to an embodiment of the present invention; Figure 3 It is a structural schematic diagram of a commodity identification device and method based on an improved convolutional neural network according to an embodiment of the present invention.

[0019] Description of reference numerals: exist Figure 1-Figure 3 In the figure, the bottom plate 1; the storage plate 2; the first track 3; the first slider 4; the first telescopic rod 5; the baffle 6; the second track 7; the connecting rod 8; the fill light 9; the second slider 10; the camera 11; the second telescopic rod 12; the spring 13; and the proximity switch 14. DETAILED DESCRIPTION

[0020] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0021] Embodiment 1: The present invention provides a commodity identification device and method based on an improved convolutional neural network, comprising: a data acquisition module, the data acquisition module comprising a bottom plate 1, a first track 3 is arranged on the bottom plate 1, a first slider 4 is arranged on the first track 3 to slide relatively, a second track 7 is fixed on the first slider 4, and a second slider 10 is arranged on the second track 7 to slide relatively. The bottom plate 1 and the storage plate 2 are both metal plates, the first track 3 and the second track 7 are both metal tracks, and the first slider 4 and the second slider 10 are both metal sliders.

[0022] A baffle plate 6 is provided at the end of the first track 3, and a first telescopic rod 5 for driving the first slider 4 to slide on the first track 3 is provided between the baffle plate 6 and the first slider 4. A second telescopic rod 12 for driving the second slider 10 to slide on the second track 7 is provided between the first slider 4 and the second slider 10. The baffle plate 6 is a metal plate, and the first telescopic rod 5 and the second telescopic rod 12 are both electric telescopic rods with a power-off self-locking function.

[0023] A camera 11 is arranged on the second slider 10, a layer of iron sheet is pasted on the back of the second track 7, a magnet is adsorbed on the iron sheet, a connecting rod 8 is connected to the magnet, and a fill light 9 is arranged on the connecting rod 8. Multiple fill lights 9 can be arranged to fill in the field of vision when the light is poor.

[0024] A plurality of springs 13 are arranged on the bottom plate 1, and the springs 13 are connected to the storage plate 2. A plurality of proximity switches 14 are arranged on the bottom plate 1, and the proximity switches 14 are arranged below the storage plate 2. When the storage plate 2 is pressed down, the proximity switches 14 will be triggered.

[0025] The identification methods include: Step 1: Place the commodity to be identified on the storage plate 2, trigger the proximity switch 14, and the camera 11 takes a photo; Step 2: Input the photos into the improved convolutional neural network model to classify and identify the products; Step 3: Output the classification results of classification recognition.

[0026] In Step 1, at least one of the proximity switches 14 is triggered to enable the camera 11 to take pictures. The triggering modes of the proximity switch 14 include a first triggering mode and a second triggering mode. The first triggering mode is triggered by the mass of the product itself when the gravity of the product itself is greater than the critical pressure of the placement plate 2 overcoming the pressure of the spring 13 to trigger the proximity switch 14; the second triggering mode is triggered by the user manually pressing the placement plate 2 when the gravity of the product itself is less than the critical pressure of the placement plate 2 overcoming the pressure of the spring 13 to trigger the proximity switch 14, and the critical pressure is 20N.

[0027] In Step 2, the improved convolutional neural network model includes: Feature extraction module: includes several convolutional layers and pooling layers to extract image information and generate feature maps; Feature transformation module: performs feature transformation on the feature map extracted by the feature extraction module; Classification module: Use a multi-layer perceptron to process the data and obtain the final classification result.

[0028] The size of the collected image is 1024x1024. This image is used as the input of the model. The classification categories include multiple categories, each category includes a product. When training the model, at least 200 images are used for each category.

[0029] The specific structure of the improved convolutional neural network model includes: Feature extraction module: Convolutional module 1: Convolutional layer: The number of convolution kernels is 2, the convolution kernel size is 3x3, the stride is 1, and the padding is 1; Activation function: RELU; Pooling layer: type is maximum pooling, pooling size is 2x2, and stride is 2; Convolutional module 2: Convolutional layer: The number of convolution kernels is 2, the convolution kernel size is 3x3, the stride is 1, and the padding is 1; Activation function: RELU; Pooling layer: type is maximum pooling, pooling size is 2x2, and stride is 2; Convolutional module 3: Convolutional layer: The number of convolution kernels is 2, the convolution kernel size is 3x3, the stride is 1, and the padding is 1; Activation function: RELU; Pooling layer: type is maximum pooling, pooling size is 2x2, and stride is 2; Convolutional module 4: Convolutional layer: The number of convolution kernels is 2, the convolution kernel size is 3x3, the stride is 1, and the padding is 1; Activation function: RELU; Pooling layer: type is maximum pooling, pooling size is 2x2, and stride is 2; Convolutional module 5: Convolutional layer: The number of convolution kernels is 2, the convolution kernel size is 3x3, the stride is 1, and the padding is 1; Activation function: RELU; Pooling layer: type is maximum pooling, pooling size is 2x2, and stride is 2; Convolutional module 6: Convolutional layer: The number of convolution kernels is 2, the convolution kernel size is 3x3, the stride is 1, and the padding is 1; Activation function: RELU; Pooling layer: type is maximum pooling, pooling size is 2x2, and stride is 4; The feature map output by feature extraction: the size is 8x8, and the number of feature maps is 64.

[0030] Feature transformation module: From each feature map output by the feature extraction module, 6 features are extracted, namely: feature 1, feature 2, feature 3, feature 4, feature 5, feature 6, where: Feature 1: Same as each feature map output by the feature extraction module, without any processing, size 8x8; Feature 2: Each feature map output by the feature extraction module is added to its own transposed matrix to obtain feature 2, with a size of 8x8; Feature 3: retain the 1st to 2nd row elements of feature 2, set all other elements to 0, and the size is 8x8; Feature 4: retain the 3rd-4th row elements of feature 2, set all other elements to 0, and the size is 8x8; Feature 5: retain the 5th to 6th row elements of feature 2, set all other elements to 0, and the size is 8x8; Feature 6: retain the 7th-8th row elements of feature 2, set all other elements to 0, and the size is 8x8; Multiply feature 1, feature 2, feature 3, feature 4, feature 5, and feature 6 with the corresponding weight coefficients (weight coefficient 1, weight coefficient 2, weight coefficient 3, weight coefficient 4, weight coefficient 5, and weight coefficient 6) to obtain the output features of the feature transformation module, where weight coefficient 1, weight coefficient 2, weight coefficient 3, weight coefficient 4, weight coefficient 5, and weight coefficient 6 are learnable parameters of the model and participate in neural network training. Their initialization values ​​are all 0.17; The features output by the feature transformation module are of size 8x8 and 64 in number.

[0031] Classification module: First, the 64 features of size 8x8 are flattened into a vector of 4096 dimensions; Linear layer 1: the number of input features is 4096, the number of output features is 32768, and there is no bias; Linear layer 2: the number of input features is 32768, the number of output features is 1024, and there is no bias; Linear layer 3: the number of input features is 1024, the number of output features is 64, and there is no bias; Linear layer 4: The number of input features is 64, the number of output features is N, and there is no bias term, where N represents the number of categories of samples used in training; As a preferred solution, the Dropout technology is used to add a Dropout layer between the fully connected layers to reduce overfitting.

[0032] The last layer is the classification layer, which uses the Softmax activation function to output the classification probability.

[0033] Learning rate: set to 0.0001, which is determined through tuning to better balance the model convergence speed and stability.

[0034] Batch size: 128 samples are processed per iteration to adapt to the limitations of video memory and computing resources.

[0035] Optimizer: The Adam optimizer is used, which can automatically adjust the learning rate and has better effects on sparse gradients and non-stationary targets.

[0036] Loss function: The cross entropy loss function is used, which is suitable for multi-classification tasks.

[0037] Number of training rounds: set to 100 rounds. Experiments have verified that this setting can enable the model to fully converge.

[0038] The ratio of training set to test set is 9:1, and 90% of the data are randomly selected for training.

[0039] For example, using 200 images of tomatoes, potatoes, bananas, oranges, instant noodles, wet wipes, toothpaste, shampoo, conditioner, and shower gel for training, the test accuracy is 99.3%.

[0040] Embodiment 2: The present invention provides a commodity identification device and method based on an improved convolutional neural network, comprising: a data acquisition module, the data acquisition module comprising a bottom plate 1, a first track 3 is arranged on the bottom plate 1, a first slider 4 is arranged on the first track 3 to slide relatively, a second track 7 is fixed on the first slider 4, and a second slider 10 is arranged on the second track 7 to slide relatively. The bottom plate 1 and the storage plate 2 are both metal plates, the first track 3 and the second track 7 are both metal tracks, and the first slider 4 and the second slider 10 are both metal sliders.

[0041] A baffle plate 6 is provided at the end of the first track 3, and a first telescopic rod 5 for driving the first slider 4 to slide on the first track 3 is provided between the baffle plate 6 and the first slider 4. A second telescopic rod 12 for driving the second slider 10 to slide on the second track 7 is provided between the first slider 4 and the second slider 10. The baffle plate 6 is a metal plate, and the first telescopic rod 5 and the second telescopic rod 12 are both electric telescopic rods with a power-off self-locking function.

[0042] A camera 11 is arranged on the second slider 10, a layer of iron sheet is pasted on the back of the second track 7, a magnet is adsorbed on the iron sheet, a connecting rod 8 is connected to the magnet, and a fill light 9 is arranged on the connecting rod 8. Multiple fill lights 9 can be arranged to fill in the field of vision when the light is poor.

[0043] A plurality of springs 13 are arranged on the bottom plate 1, and the springs 13 are connected to the storage plate 2. A plurality of proximity switches 14 are arranged on the bottom plate 1, and the proximity switches 14 are arranged below the storage plate 2. When the storage plate 2 is pressed down, the proximity switches 14 will be triggered.

[0044] The identification methods include: Step 1: Place the commodity to be identified on the storage plate 2, trigger the proximity switch 14, and the camera 11 takes a photo; Step 2: Input the photos into the improved convolutional neural network model to classify and identify the products; Step 3: Output the classification results of classification recognition.

[0045] In Step 1, at least one of the proximity switches 14 is triggered to enable the camera 11 to take pictures. The triggering modes of the proximity switch 14 include a first triggering mode and a second triggering mode. The first triggering mode is triggered by the mass of the product itself when the gravity of the product itself is greater than the critical pressure of the placement plate 2 overcoming the pressure of the spring 13 to trigger the proximity switch 14; the second triggering mode is triggered by the user manually pressing the placement plate 2 when the gravity of the product itself is less than the critical pressure of the placement plate 2 overcoming the pressure of the spring 13 to trigger the proximity switch 14, and the critical pressure is 20N.

[0046] In Step 2, the improved convolutional neural network model includes: Feature extraction module: includes several convolutional layers and pooling layers to extract image information and generate feature maps; Feature transformation module: performs feature transformation on the feature map extracted by the feature extraction module; Classification module: Use a multi-layer perceptron to process the data and obtain the final classification result.

[0047] The size of the collected image is 1024x1024. This image is used as the input of the model. The classification categories include multiple categories, each category includes a product. When training the model, at least 200 images are used for each category.

[0048] The specific structure of the improved convolutional neural network model includes: Feature extraction module: Convolutional module 1: Convolutional layer: The number of convolution kernels is 2, the convolution kernel size is 7x7, the stride is 1, and the padding is 3; Activation function: RELU; Pooling layer: type is average pooling, pooling size is 2x2, and stride is 2; Convolutional module 2: Convolutional layer: The number of convolution kernels is 2, the convolution kernel size is 7x7, the stride is 1, and the padding is 3; Activation function: RELU; Pooling layer: type is average pooling, pooling size is 2x2, and stride is 2; Convolutional module 3: Convolutional layer: The number of convolution kernels is 2, the convolution kernel size is 3x3, the stride is 1, and the padding is 1; Activation function: RELU; Pooling layer: type is maximum pooling, pooling size is 2x2, and stride is 2; Convolutional module 4: Convolutional layer: The number of convolution kernels is 2, the convolution kernel size is 3x3, the stride is 1, and the padding is 1; Activation function: RELU; Pooling layer: type is maximum pooling, pooling size is 2x2, and stride is 2; Convolutional module 5: Convolutional layer: The number of convolution kernels is 2, the convolution kernel size is 3x3, the stride is 1, and the padding is 1; Activation function: RELU; Pooling layer: type is maximum pooling, pooling size is 2x2, and stride is 2; Convolutional module 6: Convolutional layer: The number of convolution kernels is 2, the convolution kernel size is 5x5, the stride is 1, and the padding is 2; Activation function: RELU; Pooling layer: type is maximum pooling, pooling size is 4x4, and stride is 4; The feature map output by feature extraction: the size is 8x8, and the number of feature maps is 64.

[0049] Feature transformation module: From each feature map output by the feature extraction module, 6 features are extracted, namely: feature 1, feature 2, feature 3, feature 4, feature 5, feature 6, where: Feature 1: Same as each feature map output by the feature extraction module, without any processing, size 8x8; Feature 2: Each feature map output by the feature extraction module is added to its own transposed matrix to obtain feature 2, with a size of 8x8; Feature 3: retain the 1st to 2nd row elements of feature 2, set all other elements to 0, and the size is 8x8; Feature 4: retain the 3rd-4th row elements of feature 2, set all other elements to 0, and the size is 8x8; Feature 5: retain the 5th to 6th row elements of feature 2, set all other elements to 0, and the size is 8x8; Feature 6: retain the 7th-8th row elements of feature 2, set all other elements to 0, and the size is 8x8; Multiply feature 1, feature 2, feature 3, feature 4, feature 5, and feature 6 with the corresponding weight coefficients (weight coefficient 1, weight coefficient 2, weight coefficient 3, weight coefficient 4, weight coefficient 5, and weight coefficient 6) to obtain the output features of the feature transformation module, where weight coefficient 1, weight coefficient 2, weight coefficient 3, weight coefficient 4, weight coefficient 5, and weight coefficient 6 are learnable parameters of the model and participate in neural network training. Their initialization values ​​are all 0.17; The features output by the feature transformation module are of size 8x8 and 64 in number.

[0050] Classification module: First, the 64 features of size 8x8 are flattened into a vector of 4096 dimensions; Linear layer 1: the number of input features is 4096, the number of output features is 8192, and there is no bias; Linear layer 2: the number of input features is 8192, the number of output features is 1024, and there is no bias term; Linear layer 3: the number of input features is 1024, the number of output features is 128, and there is no bias term; Linear layer 4: The number of input features is 128, the number of output features is N, and there is no bias term, where N represents the number of categories of samples used in training; As a preferred solution, the Dropout technology is used to add a Dropout layer between the fully connected layers to reduce overfitting.

[0051] The last layer is the classification layer, which uses the Softmax activation function to output the classification probability.

[0052] Learning rate: set to 0.0001, which is determined through tuning to better balance the model convergence speed and stability.

[0053] Batch size: 128 samples are processed per iteration to adapt to the limitations of video memory and computing resources.

[0054] Optimizer: The Adam optimizer is used, which can automatically adjust the learning rate and has better effects on sparse gradients and non-stationary targets.

[0055] Loss function: The cross entropy loss function is used, which is suitable for multi-classification tasks.

[0056] Number of training rounds: set to 100 rounds. Experiments have verified that this setting can enable the model to fully converge.

[0057] The ratio of training set to test set is 9:1, and 90% of the data are randomly selected for training.

[0058] For example, using 200 images of tomatoes, potatoes, bananas, oranges, instant noodles, wet wipes, toothpaste, shampoo, conditioner, and shower gel for training, the test accuracy is 99.5%.

[0059] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A commodity recognition device and method based on an improved convolutional neural network, comprising: The data acquisition module is characterized in that: the data acquisition module includes a base plate, a first track is provided on the base plate, a first slider is provided on the first track for relative sliding, a second track is fixed on the first slider, and a second slider is provided on the second track for relative sliding.

2. The device and method for commodity identification based on an improved convolutional neural network according to claim 1, characterized in that: A baffle is provided at the end of the first track, and a first telescopic rod for driving the first slider to slide on the first track is provided between the baffle and the first slider.

3. The device and method for commodity identification based on an improved convolutional neural network according to claim 1, characterized in that: A second telescopic rod is provided between the first slider and the second slider for driving the second slider to slide on the second track.

4. The device and method for commodity identification based on an improved convolutional neural network according to claim 1, characterized in that: A camera is provided on the second sliding block, a layer of iron sheet is pasted on the back of the second track, a magnet is adsorbed on the iron sheet, a connecting rod is connected to the magnet, and a fill light is provided on the connecting rod.

5. The commodity identification device and method based on the improved convolutional neural network according to claim 1, characterized in that: A plurality of springs are provided on the bottom plate, and a storage plate is connected to the springs.

6. The device and method for commodity identification based on an improved convolutional neural network according to claim 5, characterized in that: A plurality of proximity switches are arranged on the bottom plate, and the proximity switches are arranged below the storage plate.

7. The device and method for commodity identification based on an improved convolutional neural network according to claim 1, characterized in that: The following identification methods are included: Step 1: Place the item to be identified on the shelf, trigger the proximity switch, and the camera takes a photo; Step 2: Input the photos into the improved convolutional neural network model to classify and identify the products; Step 3: Output the classification results of classification recognition.

8. The device and method for commodity identification based on an improved convolutional neural network according to claim 7, characterized in that: In Step 1, triggering one of the proximity switches can cause the camera to take a picture. The proximity switch triggering modes include a first triggering mode and a second triggering mode. The first triggering mode is triggered by the mass of the product itself when the weight of the product itself is greater than the critical pressure of the placement plate overcoming the spring pressure to trigger the proximity switch; the second triggering mode is triggered by the user manually pressing the placement plate when the weight of the product itself is less than the critical pressure of the placement plate overcoming the spring pressure to trigger the proximity switch. The critical pressure is 10-50N.

9. The device and method for commodity identification based on an improved convolutional neural network according to claim 7, characterized in that: In Step 2, improving the convolutional neural network model includes: Feature extraction module: includes several convolutional layers and pooling layers to extract image information and generate feature maps; Feature transformation module: performs feature transformation on the feature map extracted by the feature extraction module; Classification module: Use a multi-layer perceptron to process the data and obtain the final classification results.

Images