Image orientation recognition method and device, equipment and storage medium

Abstract

The invention discloses an image orientation recognition method, device and equipment and a storage medium, and the method comprises the steps: carrying out the preprocessing of a target image through employing a projection neural network, obtaining standard image data, and carrying out the preprocessing, including dimension reduction, noise reduction, normalization, cutting and gain control; the standard image data are coded through a preset KC layer according to APL neurons and a WTA mechanism, coded image data are obtained, and a first synaptic weight obeying random distribution is included between the projection neural network and the preset KC layer; and inputting the coded image data into a preset MBON layer for orientation analysis to obtain an orientation recognition result, a second synaptic weight updated based on a preset updating algorithm being included between the preset KC layer and the preset MBON layer. According to the method, the problem of disastrous forgetting or low learning speed caused by large distribution difference of samples of the same label in the existing image orientation recognition model is avoided, and the problem of low learning and recognition efficiency of rare detail features in the image caused by using a convolution kernel is avoided.

Description

Technical Field

[0001] The present application relates to the field of biological neural network technology, and in particular to an image orientation recognition method, device, equipment and storage medium. Background Art

[0002] In the natural environment, many cues can indicate direction, such as the direction of the sun, the offset of sunlight, the direction the wind blows, and landmarks. Animals can use them for orientation and navigation. Most of these cues are perceived through vision. Even simple insects can use visual memory to remember the way home by finding a familiar direction after walking along the route once. Their lightweight neural circuits take less time to remember directions than typical artificial neural networks (ANNs). Therefore, it is possible to retrieve directional memories obtained from visual signals by studying the structure and learning rules of these circuits.

[0003] There are already many models for finding the orientation of objects in the image plane; however, these methods usually involve deep networks and require random sorting of many data samples during training. The operations are complex and require a long training time, making it difficult to adapt to the requirements of real-time, fast, and low-power learning in application scenarios such as robotics. Summary of the invention

[0004] The present application provides an image orientation recognition method, device, equipment and storage medium, which are used to solve the technical problem that the existing image orientation recognition model network is relatively complex and has high requirements for training samples, so the operation is complicated and requires a long training time, resulting in low recognition efficiency.

[0005] In view of this, the first aspect of the present application provides an image orientation recognition method, comprising:

[0006] Preprocessing the target image using a projected neural network to obtain standard image data, wherein the preprocessing includes dimensionality reduction, noise reduction, normalization, cropping, and gain control;

[0007] The standard image data is encoded by a preset KC layer according to the APL neuron and the WTA mechanism to obtain encoded image data, wherein the projection neural network and the preset KC layer include a first synaptic weight that obeys a random distribution;

[0008] The encoded image data is input into a preset MBON layer for orientation analysis to obtain an orientation recognition result, and a second synaptic weight updated based on a preset update algorithm is included between the preset KC layer and the preset MBON layer.

[0009] Preferably, encoding the standard image data according to the APL neuron and the WTA mechanism through the preset KC layer to obtain the encoded image data includes:

[0010] Randomly generate a first synaptic weight in binary form between the projection neural network and the preset KC layer based on Bernoulli distribution;

[0011] Performing linear mapping according to the first synaptic weight to obtain mapping data;

[0012] The mapping data is laterally inhibited according to the APL neuron and the WTA mechanism, and encoded into a sparse vector to obtain encoded image data.

[0013] Preferably, the step of inputting the encoded image data into a preset MBON layer for orientation analysis to obtain an orientation recognition result includes:

[0014] Based on a preset update algorithm, the synaptic weight between the preset KC layer and the preset MBON layer is updated and calculated to obtain a second synaptic weight;

[0015] A mapping-based orientation activity analysis is performed on the encoded image data according to the second synaptic weight to obtain an orientation recognition result.

[0016] Preferably, the updating and calculating the synaptic weight between the preset KC layer and the preset MBON layer based on a preset updating algorithm to obtain the second synaptic weight includes:

[0017] The preset update algorithm is an update algorithm based on the Hebbian rule;

[0018] Based on the Hebbian rule, the synaptic weights between the preset KC layer and the preset MBON layer are learned and updated to obtain a second synaptic weight under convergence.

[0019] Preferably, the updating and calculating the synaptic weight between the preset KC layer and the preset MBON layer based on a preset updating algorithm to obtain the second synaptic weight includes:

[0020] The preset update algorithm is a binary-based update algorithm;

[0021] Based on the binary system, the synaptic weight between the preset KC layer and the preset MBON layer is updated in binary mode in real time to obtain a second synaptic weight.

[0022] A second aspect of the present application provides an image orientation recognition device, comprising:

[0023] A preprocessing unit, used for preprocessing the target image using a projected neural network to obtain standard image data, wherein the preprocessing includes dimensionality reduction, noise reduction, normalization, cropping and gain control;

[0024] An image encoding unit, used for encoding the standard image data according to the APL neuron and the WTA mechanism through a preset KC layer to obtain encoded image data, wherein the projection neural network and the preset KC layer include first synaptic weights that obey random distribution;

[0025] The orientation recognition unit is used to input the encoded image data into a preset MBON layer for orientation analysis to obtain an orientation recognition result, wherein the preset KC layer and the preset MBON layer include a second synaptic weight updated based on a preset update algorithm.

[0026] Preferably, the image encoding unit is specifically used for:

[0027] Randomly generate a first synaptic weight in binary form between the projection neural network and the preset KC layer based on Bernoulli distribution;

[0028] Performing linear mapping according to the first synaptic weight to obtain mapping data;

[0029] The mapping data is laterally inhibited according to the APL neuron and the WTA mechanism, and encoded into a sparse vector to obtain encoded image data.

[0030] Preferably, the orientation identification unit is specifically used for:

[0031] Based on a preset update algorithm, the synaptic weight between the preset KC layer and the preset MBON layer is updated and calculated to obtain a second synaptic weight;

[0032] A mapping-based orientation activity analysis is performed on the encoded image data according to the second synaptic weight to obtain an orientation recognition result.

[0033] A third aspect of the present application provides an image orientation recognition device, the device comprising a processor and a memory;

[0034] The memory is used to store program code and transmit the program code to the processor;

[0035] The processor is used to execute the image orientation recognition method of the first aspect according to instructions in the program code.

[0036] A fourth aspect of the present application provides a computer-readable storage medium, which is used to store program code, and the program code is used to execute the image orientation recognition method of the first aspect.

[0037] It can be seen from the above technical solutions that the embodiments of the present application have the following advantages:

[0038] In the present application, an image orientation recognition method is provided, including: using a projection neural network to preprocess a target image to obtain standard image data, the preprocessing includes dimensionality reduction, noise reduction, normalization and gain control; encoding the standard image data according to APL neurons and WTA mechanism through a preset KC layer to obtain encoded image data, and the projection neural network and the preset KC layer include a first synaptic weight that obeys a random distribution; inputting the encoded image data into a preset MBON layer for orientation analysis to obtain an orientation recognition result, and the preset KC layer and the preset MBON layer include a second synaptic weight that is updated based on a preset update algorithm.

[0039] The image orientation recognition method provided by the present application analyzes and recognizes the orientation of the target in the image based on the mushroom body model, and incrementally learns the orientation of the image object according to the biological learning rules. The model in this process only needs one observation to learn the direction of the target in the image and obtain the orientation recognition result; and the network structure is relatively simple, and the requirements for training time and training samples are low. It can quickly and accurately obtain the orientation recognition result, improve the recognition efficiency, and meet the needs of various application scenarios. Therefore, the present application can solve the technical problem that the existing image orientation recognition model network is relatively complex and has high requirements for training samples, so the operation is complicated and requires a long training time, resulting in low recognition efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 A schematic diagram of a flow chart of an image orientation recognition method provided in an embodiment of the present application;

[0041] Figure 2 A schematic diagram of the structure of an image orientation recognition device provided in an embodiment of the present application;

[0042] Figure 3 A schematic diagram of the biological structure of the Drosophila mushroom body and the corresponding simplified computational model network structure provided in the embodiments of the present application;

[0043] Figure 4 An example of the object orientation of the COIL-100 dataset provided for the experimental application example of this application;

[0044] Figure 5 Schematic diagram of the accuracy and loss rate curves on the COIL-100 dataset provided for the experimental application example of this application. DETAILED DESCRIPTION

[0045] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0046] It is very interesting to explore how insects can remember their direction by observing a landmark once without caching image data, and use its underlying mechanism to retrieve or detect the direction of objects. The neural circuits of insects are able to perform sparse encoding without learning, and the resulting sparse expression reduces conflicts in learning. The mushroom body (MB) of Drosophila is a typical example; the field of biology has conducted detailed research on the nervous system of Drosophila, reconstructed its circuits in three dimensions, and analyzed its connection group data. Using these biological data, the working mechanism of Drosophila can be explained more accurately. Based on the working principle of the Drosophila mushroom body, a more efficient and reliable image orientation recognition method can be conceived, and this application is a study based on this principle.

[0047] For navigation tasks using visual information collected near the ground, landmark objects usually have a fixed angle relative to the ground, so the direction in the view plane is not that important relative to the angle of the object's direction to the observer. Traditional object angle recognition focuses on determining where the object is located in a certain view to the observer, but the model of this application focuses on the orientation of the object itself in the image.

[0048] For easier understanding, see Figure 1 , an embodiment of an image orientation recognition method provided by the present application includes:

[0049] Step 101: Use a projection neural network to preprocess the target image to obtain standard image data, and the preprocessing includes dimensionality reduction, noise reduction, normalization, cropping and gain control.

[0050] It should be noted that this embodiment designs an image target orientation recognition model based on mushroom bodies, and the model specifically includes three layers; the first layer is a projection neural network layer, which is composed of projection neurons (PNs) and is used to preprocess images; the second layer is a preset KC layer, which includes multiple Kenyon cells (KCs) and is used to map and encode images; the third layer is a preset MBON layer, which is used to learn and analyze the possibility of the orientation of the target. Therefore, before the insect receptor input is transmitted to the Kenyon cells (KCs), the projection neurons PNs in the projection neural network will perform a series of preprocessing operations on the target image. This is the design based on the processing principle of the fruit fly mushroom body (MB).

[0051] Each Kenyon cell (KC) in the Drosophila mushroom body (MB) receives signals from several specific classes of projection neurons (PNs), usually 3 to 5, and the KC population encodes the signals in a sparse coding manner. Only a small fraction of KCs can be activated simultaneously, usually 5% to 15%, because the anterior paired lateral neuron APL receives input from all KCs and inhibits them. The axons of KCs traverse multiple segments, each of which usually has one mushroom body output neuron (MBON) for outputting direction unfamiliarity based on visual analysis, and one dopamine neuron for reward and learning regulation. Because the number of KCs is much larger than that of PNs and the activity of KCs is sparse, the existing connections between KCs and MBONs are unlikely to be affected by subsequent learning.

[0052] This is the working principle of the fruit fly mushroom body model in this embodiment. For details on the biological structure and the simplified mushroom body model in this embodiment, please refer to Figure 3 The biological structure of the mushroom body shows plastic connections from olfactory receptor neurons to mushroom body output neurons, while the simplified models describe the mapping relationship between neurons.

[0053] It can be seen that the mushroom body model of the present application includes quasi-projection neurons PNs, a preset KC layer for encoding, and a preset MBON layer for output. The model ignores the morphology of neurons and uses a firing-rate neuron model instead of a spiking neuron model; it also ignores the dynamics inside neurons and regards the synapses between neurons as linear mappings. In this way, the complexity of the model is simplified to the greatest extent, and the processing efficiency of the model is improved.

[0054] Although the actual preprocessing of visual signals by insects may be complex, according to previous models, mushroom structures can be processed and learned on images without complex feature extraction, only downsampling and normalization. In the first layer of the model of this embodiment, the original image is normalized so that the average pixel intensity is 0. Although an image has two dimensions, in order to generate this model on other modalities, each image is flattened to one dimension. Therefore, if a given dataset , Each row in represents a data point ,in is the dimension of the data point. Before the data is passed to the PNs, they are The mean Normalization. The standard image data obtained by preprocessing is expressed as:

[0055]

[0056] in, That is the standard image data, and n represents the total amount of data.

[0057] Step 102: Encode the standard image data according to the APL neuron and the WTA mechanism through the preset KC layer to obtain the encoded image data, and the projection neural network and the preset KC layer include a first synaptic weight that obeys a random distribution.

[0058] Further, step 102 includes:

[0059] Randomly generate a first synaptic weight in binary form between the projection neural network and the preset KC layer based on Bernoulli distribution;

[0060] Performing linear mapping according to the first synaptic weight to obtain mapping data;

[0061] The mapping data is laterally inhibited according to the APL neuron and the WTA mechanism, and encoded into a sparse vector to obtain encoded image data.

[0062] It should be noted that the projection neural network and the preset KC layer perform mapping encoding operations on the standard image data based on the first synaptic weight that obeys the random distribution. The specific random distribution can be Bernoulli distribution or other distributions. This is only an example and is not limited. If the first synaptic weight is expressed as ,in, is the number of Kenyon cells (KC), the first synaptic weight The elements of are random and binary and follow a Bernoulli distribution, i.e. ,in, represents the probability of connection, b represents the expectation that the projection neurons PNs are connected to the Kenyon cells, then the mapping based on the first synaptic weight can be expressed as:

[0063]

[0064] in, This is the standard image data obtained above. It is mapping data.

[0065] In the mapping process of MB, APL neurons will produce lateral inhibition on KCs, so only the most active KC can play a role. This WTA inhibitory coding mechanism is approximated by keeping the output values ​​of the first h most excited KCs, and setting the others to zero. The specific process is expressed as:

[0066]

[0067] Among them, h can control the encoding wash property, and j is the active and excited KC index. After WTA suppression encoding, the output of KCs is expressed as .

[0068] Since a KC that is always active provides less useful information, a threshold can be used to disable such a KC. For example, if the threshold is set to 0.25, if the KC remains active in more than a quarter of the images, its output is 0.

[0069] Step 103: Input the encoded image data into a preset MBON layer for orientation analysis to obtain an orientation recognition result. The preset KC layer and the preset MBON layer include a second synaptic weight updated based on a preset update algorithm.

[0070] Furthermore, step 103 includes:

[0071] Based on a preset update algorithm, the synaptic weight between the preset KC layer and the preset MBON layer is updated and calculated to obtain a second synaptic weight;

[0072] A mapping-based orientation activity analysis is performed on the encoded image data according to the second synaptic weight to obtain an orientation recognition result.

[0073] Further, the synaptic weight between the preset KC layer and the preset MBON layer is updated and calculated based on a preset update algorithm to obtain a second synaptic weight, including:

[0074] The default update algorithm is the update algorithm based on the Hebbian rule;

[0075] The synaptic weights between the preset KC layer and the preset MBON layer are learned and updated based on the Hebbian rule to obtain the second synaptic weights under convergence.

[0076] Further, the synaptic weight between the preset KC layer and the preset MBON layer is updated and calculated based on a preset update algorithm to obtain a second synaptic weight, including:

[0077] The default update algorithm is a binary-based update algorithm;

[0078] The synaptic weights between the preset KC layer and the preset MBON layer are updated in binary format in real time based on the binary system to obtain the second synaptic weights.

[0079] It should be noted that in order to interpret the associative learning of insects as supervised learning in machine learning, learning can be improved by changing the weights between neural layers. Therefore, this embodiment proposes a preset update algorithm to update the synaptic weights from the preset KC layer to the preset MBON layer to obtain a more reliable MBON update weight, that is, the second synaptic weight.

[0080] If the second synaptic weight is expressed as , then different preset update algorithms can be used to obtain different synaptic weights based on different methods. The output of the mapping-based orientation activity analysis of the encoded image data according to the second synaptic weight can be expressed as:

[0081]

[0082] The activity and excitement level of each MBON indicates the possibility of the orientation angle corresponding to the sample, and the most active angle is the orientation recognition result of the target.

[0083] This embodiment uses two different methods to update and calculate the second synaptic weight. The first method is to use a method similar to the Hebbian rule to implement the plasticity rule of neurons for achieving convergence. The synaptic weight can be adjusted multiple times. The second synaptic weight is expressed as:

[0084]

[0085] in, is the learning rate. After each iterative sample, the learning rate decays according to the rule. The decay rule is expressed as: .

[0086] The second method is to update the second synaptic weight based on a binary method, which is specifically expressed as:

[0087]

[0088] This method initializes all weights between KCs and MBONs to 0. During learning, when an image is sent to the network, the most active KCs are activated, and the labels corresponding to the weights between KCs and MBONs are set to 1. Any of the above two methods can be used to update the weights to obtain the MBON update weights, i.e., the second synaptic weights.

[0089] For ease of understanding, this application conducts experiments on image orientation recognition methods on the COIL dataset, which contains 100 objects captured at 72 different angles and a total of 7200 images. The size of the original image is 128×128, and for each image, there are 128×128×3=49152 value channels because the image is RGB color. The dataset is divided into two groups according to whether the object has axial symmetry and whether it has a texture pattern, namely COIL-100-Ordinary (COIL-100-O) and COIL-100-Axisymmetric (COIL-100-AS). For COIL-100-O, the object does not have axial symmetry or has a clear texture pattern. For COIL-100-AS, the object is axially symmetric and has no obvious texture pattern. In COIL-100-AS, different perspectives of the same object are so similar that the human eye cannot even distinguish them. See Figure 4 View of two objects, the first row is from COIL-100-O and the second row is from COIL-100-AS.

[0090] In this study, the orientation of the object is discretized into 72 orientations ranging from 0 to 355, and k is set to 72. The accuracy of the model provided in this application in distinguishing objects with trained orientations and unknown orientations is evaluated.

[0091] The first experiment on the COIL dataset is the analysis of object orientation. This experiment does not distinguish between training and test sets. Instead, the model should analyze the orientation of objects in previously viewed images. This is conceptually simple, but since the same objects share the same features, it will cause interference when typical artificial neural networks learn orientation. The impact of the number of active KCs on retrieval accuracy is evaluated. As the number of KCs increases, the accuracy of the proposed method on both datasets increases, and the accuracy tends to stabilize when the number of KCs approaches 10,000. For details, please refer to Figure 5 (a). This application selected 10240 MBONs. The results show that the algorithm model of this application can achieve an accuracy of more than 90% on two data sets after learning once, thereby identifying the direction of the observed object.

[0092] The baseline model requires longer training time for the same performance. The accuracy, training time, and incremental learning ability of the two weight update methods of the Benshin algorithm model are compared with those of deep convolutional neural networks (such as AlexNet, GoogleNet, VGG16, and ResNet50). The accuracy and loss of the baseline model change with the increase in the number of iterations. Figure 5 As shown. Figure 5In (b), the methods of the present application are shown as horizontal lines because they only need to be learned once. Other models take 1.6 to 80.6 times longer to achieve similar performance. Moreover, the methods of the present application have not accelerated the model on the GPU in the experiment. Therefore, the experimental results can confirm that the present application can still accurately and reliably identify the orientation of the target object in the image without increasing the complexity of the model and model training, and has strong comprehensive performance.

[0093] The incremental learning capability calculates the accuracy of previously trained objects to evaluate whether the model can correctly identify the direction of previous objects after acquiring new objects. More precisely, after training the i-th object, the model is evaluated for the accuracy of the direction of each object ≤i angles. It can be found that the model of the present application can remember old memories while learning new knowledge, even for axisymmetric objects that are difficult for humans to recognize. All baseline models will forget the learned objects after 10 optimization iterations, but the model of the present application is almost unaffected by sequential samples, that is, there is no need to sort the samples at all during the model training process, which can minimize the complexity of model training and improve the efficiency of model training.

[0094] The image orientation recognition method provided in the embodiment of the present application analyzes and recognizes the orientation of the target in the image based on the mushroom body model, and incrementally learns the orientation of the image object according to the biological learning rules. The model in this process only needs one observation to learn the direction of the target in the image and obtain the orientation recognition result; and the network structure is relatively simple, with low requirements for training time and training samples, and can quickly and accurately obtain the orientation recognition result, improve recognition efficiency, and meet the needs of various application scenarios. Therefore, the embodiment of the present application can solve the technical problem that the existing image orientation recognition model network is relatively complex and has high requirements for training samples, so the operation is complicated and requires a long training time, resulting in low recognition efficiency.

[0095] For easier understanding, see Figure 2 , the present application provides an embodiment of an image orientation recognition device, including:

[0096] A preprocessing unit 201 is used to preprocess the target image using a projected neural network to obtain standard image data, wherein the preprocessing includes dimensionality reduction, noise reduction, normalization, cropping and gain control;

[0097] An image encoding unit 202, configured to encode the standard image data according to the APL neuron and the WTA mechanism through a preset KC layer to obtain encoded image data, wherein the projection neural network and the preset KC layer include first synaptic weights that obey a random distribution;

[0098] The orientation recognition unit 203 is used to input the encoded image data into the preset MBON layer for orientation analysis to obtain an orientation recognition result. The preset KC layer and the preset MBON layer include a second synaptic weight updated based on a preset update algorithm.

[0099] Furthermore, the image encoding unit 202 is specifically configured to:

[0100] Randomly generate a first synaptic weight in binary form between the projection neural network and the preset KC layer based on Bernoulli distribution;

[0101] Performing linear mapping according to the first synaptic weight to obtain mapping data;

[0102] The mapping data is laterally inhibited according to the APL neuron and the WTA mechanism, and encoded into a sparse vector to obtain encoded image data.

[0103] Further, the direction identification unit 203 is specifically configured to:

[0104] Based on a preset update algorithm, the synaptic weight between the preset KC layer and the preset MBON layer is updated and calculated to obtain a second synaptic weight;

[0105] A mapping-based orientation activity analysis is performed on the encoded image data according to the second synaptic weight to obtain an orientation recognition result.

[0106] The present application also provides an image orientation recognition device, the device comprising a processor and a memory;

[0107] The memory is used to store the program code and transmit the program code to the processor;

[0108] The processor is used to execute the image orientation recognition method in the above method embodiment according to the instructions in the program code.

[0109] The present application also provides a computer-readable storage medium, which is used to store program codes, and the program codes are used to execute the image orientation recognition method in the above method embodiment.

[0110] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0111] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0112] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0113] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for executing all or part of the steps of the method described in each embodiment of the present application through a computer device (which can be a personal computer, server, or network device, etc.). The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (full name in English: Read-Only Memory, English abbreviation: ROM), random access memory (full name in English: Random Access Memory, English abbreviation: RAM), disk or optical disk and other media that can store program code.

[0114] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for image orientation recognition, characterized in that: include: Preprocessing the target image using a projected neural network to obtain standard image data, wherein the preprocessing includes dimensionality reduction, noise reduction, normalization, cropping, and gain control; The standard image data is encoded by a preset KC layer according to the APL neuron and the WTA mechanism to obtain encoded image data, wherein the projection neural network and the preset KC layer include a first synaptic weight that obeys a random distribution; The encoded image data is input into a preset MBON layer for orientation analysis to obtain an orientation recognition result, and a second synaptic weight updated based on a preset update algorithm is included between the preset KC layer and the preset MBON layer.

2. The image orientation recognition method according to claim 1, characterized in that: The method of encoding the standard image data according to the APL neuron and the WTA mechanism through the preset KC layer to obtain the encoded image data includes: Randomly generate a first synaptic weight in binary form between the projection neural network and the preset KC layer based on Bernoulli distribution; Performing linear mapping according to the first synaptic weight to obtain mapping data; The mapping data is laterally inhibited according to the APL neuron and the WTA mechanism, and encoded into a sparse vector to obtain encoded image data.

3. The image orientation recognition method according to claim 1, characterized in that: The step of inputting the encoded image data into a preset MBON layer for orientation analysis to obtain an orientation recognition result comprises: Based on a preset update algorithm, the synaptic weight between the preset KC layer and the preset MBON layer is updated and calculated to obtain a second synaptic weight; A mapping-based orientation activity analysis is performed on the encoded image data according to the second synaptic weight to obtain an orientation recognition result.

4. The image orientation recognition method according to claim 3, characterized in that: The updating and calculating the synaptic weight between the preset KC layer and the preset MBON layer based on a preset updating algorithm to obtain a second synaptic weight includes: The preset update algorithm is an update algorithm based on the Hebbian rule; Based on the Hebbian rule, the synaptic weights between the preset KC layer and the preset MBON layer are learned and updated to obtain a second synaptic weight under convergence.

5. The image orientation recognition method according to claim 3, characterized in that: The updating and calculating the synaptic weight between the preset KC layer and the preset MBON layer based on a preset updating algorithm to obtain a second synaptic weight includes: The preset update algorithm is a binary-based update algorithm; Based on the binary system, the synaptic weight between the preset KC layer and the preset MBON layer is updated in binary mode in real time to obtain a second synaptic weight.

6. An image orientation recognition device, characterized in that: include: A preprocessing unit, used for preprocessing the target image using a projected neural network to obtain standard image data, wherein the preprocessing includes dimensionality reduction, noise reduction, normalization, cropping and gain control; An image encoding unit, used for encoding the standard image data according to the APL neuron and the WTA mechanism through a preset KC layer to obtain encoded image data, wherein the projection neural network and the preset KC layer include first synaptic weights that obey random distribution; The orientation recognition unit is used to input the encoded image data into a preset MBON layer for orientation analysis to obtain an orientation recognition result, wherein the preset KC layer and the preset MBON layer include a second synaptic weight updated based on a preset update algorithm.

7. The image orientation recognition device according to claim 6, characterized in that: The image encoding unit is specifically used for: Randomly generate a first synaptic weight in binary form between the projection neural network and the preset KC layer based on Bernoulli distribution; Performing linear mapping according to the first synaptic weight to obtain mapping data; The mapping data is laterally inhibited according to the APL neuron and the WTA mechanism, and encoded into a sparse vector to obtain encoded image data.

8. The image orientation recognition device according to claim 6, characterized in that: The orientation recognition unit is specifically used for: Based on a preset update algorithm, the synaptic weight between the preset KC layer and the preset MBON layer is updated and calculated to obtain a second synaptic weight; A mapping-based orientation activity analysis is performed on the encoded image data according to the second synaptic weight to obtain an orientation recognition result.

9. An image orientation recognition device, characterized in that: The device comprises a processor and a memory; The memory is used to store program codes and transmit the program codes to the processor; The processor is used to execute the image orientation recognition method according to any one of claims 1 to 5 according to the instructions in the program code.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium is used to store program codes, and the program codes are used to execute the image orientation recognition method according to any one of claims 1 to 5.

Images