1172results about How to "Reduce the amount of parameters" patented technology

The invention discloses a remote sensing image semantic segmentation method based on attention multi-scale feature fusion, and the method comprises the steps: establishing a semantic segmentation network based on attention multi-scale feature fusion, building a training data set, and carrying out the network parameter training through employing the training data set. And performing semantic segmentation on to-be-tested data by using the trained network during testing. And the network is a lightweight encoder-decoder structure. Wherein the idea of an image cascade network is introduced; meanwhile, coding features and decoding features are optimized by using an attention mechanism, a multi-scale attention optimization module, a multi-scale feature fusion module and a boundary enhancement module are constructed, feature maps of different scales are extracted and fused, training is guided by using multi-scale semantic tags and boundary tags, and semantic segmentation of remote sensing images can be effectively carried out.

The invention discloses a multi-scale image semantic segmentation method. The method comprises the following steps: obtaining a to-be-segmented image and a corresponding label; constructing a full convolutional deep neural network, wherein the full convolutional deep neural network comprises a convolution module, a hole convolution module, a pyramid pooling module, a 1 * 1 * depth convolution layer and a deconvolution structure; setting hole convolution as channel-by-channel operation, and utilizing low-scale, medium-scale and high-scale characteristics in a targeted mode; training the full convolutional deep neural network, establishing a loss function, and determining parameters of the full convolutional deep neural network by training the sample image; and inputting the to-be-segmentedimage into the trained full convolutional deep neural network to obtain a semantic segmentation result. By means of the method, the image semantic segmentation problem with complex details, holes andlarge targets can be well solved while the calculated amount and the parameter number are reduced, and the consistency of category labels can be reserved while the target edges can be well segmented.

The invention discloses a full convolution neural network (FCN)-based monocular image depth estimation method. The method comprises the steps of acquiring training image data; inputting the training image data into a full convolution neural network (FCN), and sequentially outputting through pooling layers to obtain a characteristic image; subjecting each characteristic image outputted by a last pooling layer sequentially to amplification treatment to obtain a new characteristic image the same with the dimension of a characteristic image outputted by a previous pooling layer, and fusing the twocharacteristic images; sequentially fusing the outputted characteristic image of each pooling layer from back to front so as to obtain a final prediction depth image; training the parameters of the full convolution neural network (FCN) by utilizing a random gradient descent method (SGD) during training; acquiring an RGB image required for depth prediction, and inputting the RGB image into the well trained full convolution neural network (FCN) so as to obtain a corresponding prediction depth image. According to the method, the problem that the resolution of an output image is low in the convolution process can be solved. By adopting the form of the full convolution neural network, a full-connection layer is removed. The number of parameters in the network is effectively reduced.

A lightweight semantic segmentation method for a high-resolution remote sensing image comprises the steps of network construction, training and testing. Specifically, a deep semantic segmentation network of an encoder-decoder structure is constructed for a pytorch deep learning framework, after network training is carried out based on a remote sensing image data sample set, a to-be-tested remote sensing image serves as network input. A segmentation result of the remote sensing image is obtained. According to the method, on one hand, model parameters are reduced by decomposing depth separable convolution, the calculation complexity is reduced, the semantic segmentation time of the high-resolution remote sensing image is shortened, and the semantic segmentation efficiency of the high-resolution remote sensing image is improved; and on the other hand, semantic segmentation precision is improved through multi-scale feature aggregation, a spatial attention module and gating convolution, sothat the proposed lightweight deep semantic segmentation network can accurately and efficiently realize semantic segmentation of a high-resolution remote sensing image.

The invention discloses a face multi-area fusion expression recognition method based on depth learning, which comprises the following steps of detecting a face position with a detection model; obtaining the coordinates of the key points by using the key point model; aligning the eyes according to the key points of the eyes, then aligning the face according to the coordinates of the key points of the whole face, and clipping the face region by affine transformation; cutting the eye and mouth areas of the image to a certain proportion; dividing the convolution neural network into one backbone network and two branch networks; carrying out the feature fusion in the last convolution layer, and finally obtaining the expression classification results by the classifier. The method of the inventionutilizes the priori information, besides the whole face, the eyes and mouth regions are also used as the input of the network, and the network can learn the whole semantic features of facial expressions and the local features of facial expressions through model fusion, so that the method simplifies the difficulty of facial expression recognition, reduces the external noise, and has strong robustness, high accuracy, low complexity of the algorithm and so on.

The invention discloses a traffic image multi-type vehicle detection method based on deep study. The traffic image multi-type vehicle detection method based on deep study firstly combines a nerve network characteristic with a region generation algorithm, realizes two processes-region generation and region determination-at the same time through using a nerve network conventional layer, uses a background model to determine a motion area of a discrete series image targeting a specific scene and provides an extra reference basis to region generation, and combines with a vehicle detection result to perform upgrading and correction on the background model according to different conditions. Besides, the traffic image multi-type vehicle detection method also brings forward a network model compression scheme to reduce model parameters and calculation time, brings forward a new detection result optimization means based on grouped error calculation to replace a conventional non-maximum-value inhibition scheme, and improves overall detection accuracy.

The invention provides a medical image segmentation method based on an improved convolutional neural network, and belongs to the field of image processing. According to the method, an improved convolutional neural network is constructed, and a multi-channel fusion module, a multi-channel dense connection module, a down-sampling module and an up-sampling module are designed on the basis of an original framework of a standard U-Net network. The depth of the network is increased while the network training calculation amount is controlled, and redundant calculation is reduced. Gradient disappearance is avoided by improving the internal structure of the network while the network is deepened. The experimental result is that the segmentation DICE coefficient of the output prediction graph of thetrained neural network is 98.57%, the DICE coefficient segmented by the original network is 98.26%, and the effectiveness of the method is reflected.

The invention provides a voice wake-up method and device based on artificial intelligence, and computer equipment. The method comprises steps that windowing enframing operation of online recorded voice is carried out to acquire at least one voice frame; characteristic extraction of the voice frame is carried out to acquire voice characteristics; a voice wake-up model based on the convolutional neural network is utilized to carry out calculation of the static voice characteristics contained in the voice frame, and posterior probabilities of the static voice characteristics contained in the voice frame corresponding to a non-wake-up word class and a wake-up word class are acquired; if the posterior probability of the static voice characteristics contained in the voice frame corresponding to the wake-up word class is greater than a pre-determined threshold, the online recorded voice is determined to contain wake-up words. The method is advantaged in that parameter quantities of the voice wake-up model based on the convolutional neural model can be effectively reduced, so the operand reduction effect is realized, and the voice wake-up effect can be improved.

The invention discloses a fast pedestrian detection method and device. The method comprises the following steps: step S1, constructing a configurable deep model based on a convolutional neural network, and utilizing training samples to obtain parameters of a constructed network by learning to obtain a model used for a test process; and step S2, inputting test samples, utilizing a variation law ofneural-network perception domains, using different intermediate layers to detect target objects in different scale ranges through the trained model, and obtaining box graphs of the target objects in images by prediction. The method uses the different intermediate layers to detect the target objects in the certain scale ranges through utilizing the variation law of the neural-network perception domains, better adapts to relationships of the perception domains and object sizes, and effectively improves a detection result.

The invention discloses a method for detecting and identifying continuous segmented texts in an image based on SegLink and Attention-based CRNN fusion processing, and belongs to the technical field ofoptical character recognition, aiming to solve the problems of low text detection accuracy, particularly low inclined text detection accuracy, difficulty in positioning, difficulty in font segmentation and low recognition accuracy in OCR information document digitization. The method includes the steps: establishing a SegLink + CRNN model based on a Tensorflow deep learning framework, detecting text lines in an image through a SegLink network; segmenting the segmented text according to lines; extracting single-line text features through a densely connected convolutional neural network; processing the sequence information of the context in the text by the bidirectional recurrent neural network, and adopting the CTC decoding algorithm to avoid the problem of single word segmentation, and eliminate the influence of the single word segmentation link on the recognition accuracy; and further fusing an Attention mechanism during CTC transcription to improve the recognition accuracy for the text sequence characteristics. The method is applicable to printed form and handwritten form recognition, and can be applied to recognition of multilingual texts such as English and Chinese.

The invention discloses an efficient super-resolution method based on a deep back projection network. The method comprises the following steps: (1) obtaining a total training set and a test set; (2) preprocessing the total training set to complete data enhancement; (3) scaling the images in the total training set at different scales; and (4) image super-resolution reconstruction is realized basedon the convolutional neural network, and the convolutional neural network totally comprises 27 convolutional layers and specifically comprises three parts of feature extraction, error back projectionand image reconstruction. The combination of the group convolution and the 1 * 1 convolution is used for replacing the traditional convolution to redesign the iteration submodule, and the strategy caneffectively reduce the model parameter quantity and improve the model efficiency; Each iteration sub-module comprises an error feedback mechanism, so that error correction can be carried out in time;in addition, a channel weighting module is introduced, so that the model efficiency can be further improved.

The invention relates to an embedded platform real-time tumble detection method based on an improved attitude estimation algorithm. According to the method, an attitude estimation network is established by using depth separable convolution, an attention mechanism and an inverse residual error structure and is used for fall detection. The precision of the attitude estimation network is further improved; the parameter quantity and the calculation quantity are greatly reduced; the distance of each articulation point of the human body between different video frames is calculated to track the humanbody; acceleration of human body articulation points is calculated by using front and rear video frames, and whether falling occurs is judged according to the acceleration, the relative positions ofthe articulation points and the like, so that the attitude estimation network is more suitable for being deployed on an embedded platform, and a real-time effect can be achieved by deploying on a TX2embedded platform. According to the method, human body tracking is carried out by using multi-person human body joint point coordinates and skeleton information obtained in front and back frames, posture estimation is more stable due to multi-person tracking, and the problem of fall detection in a multi-person scene can be better solved.

The invention discloses a garbage can capable of automatic classifying based on visual identification and a classifying method. The garbage can comprises a garbage throwing opening, a first photoelectric switch sensor, an identifying and classifying tray, second photoelectric switch sensors, sub garbage cans, an image identification component, an STM32 controller, a double-path stepping motor driver and a garbage can shell, wherein the garbage throwing opening is formed in the side wall of the garbage can shell; the first photoelectric switch sensor is arranged in the garbage throwing opening;a plurality of sub garbage cans are arranged in the garbage can shell; a second photoelectric switch sensor is mounted in a can opening position of each sub garbage can; the identifying and classifying tray is arranged in the garbage can shell and is located above the sub garbage cans; the identifying and classifying tray comprises a garbage tray, a V-shaped baffle stepping motor, a camera, a V-shaped baffle, a support frame, a rotary baffle stepping motor and a rotary baffle. According to the garbage can, the camera is used for collecting garbage images; a TensorFlow deep learning frameworkis adopted; through transfer training of a model, the accuracy rate of garbage identification is increased.

The invention discloses a feature pyramid-based remote-sensing image time-sensitive target recognition system and method. The system comprises a target feature extraction sub-network, a feature layersub-network, a candidate area generation sub-network and a classification and regression sub-network, wherein the target feature extraction sub-network is used for carrying out multiple layers of convolution processing on a to-be-processed image and outputting the convolution processing result of each layer as a feature layer; the feature layer sub-network is used for overlapped the last feature layer and the current feature layer to obtain the current fused feature layer, wherein the topmost fused feature layer is a topmost feature layer; the candidate area generation sub-network is used forextracting candidate areas from different layers of fused feature layers; and the classification and regression sub-network is used for mapping the candidate areas to different layers of fused featurelayers so as to obtain a plurality of mapped fused feature layers, and carrying out target judgement on the plurality of mapped fused feature layers so as to output a result. According to the systemand method, the hierarchical structures of feature pyramids are utilized to ensure that all the scales of features have rich semantic information.

The invention provides an online classroom atmosphere evaluation system and method. The evaluation system comprises a video stream collection module, a data stream processing module, an image analysismodule, a classroom attendance analysis module, a classroom atmosphere evaluation module, a classroom atmosphere scoring module and a display module; . The invention collects classroom video stream data through a camera, and the captured video is captured frame by frame to obtain images; all the faces images are segmented and numbered sequentially, at the same time, the eigenvalues are assigned to the faces with the corresponding numbers, and then face recognition and facial expression recognition are carried out according to the number, so as to identify the number, emotion and movement posture of students in the video stream data; the score 0 is given to the students lowering heads in the image and the score 1 is given to students having interaction; according to the emotional analysisstrategy of students listening to the teacher, the current student listening student score is obtained and finally comprehensive evaluation of the classroom atmosphere score is obtained. The inventioncan evaluate the classroom quality on-line and in real time, and can effectively improve the evaluation effect.

The invention discloses a face image restoration method based on a multi-discriminator generative adversarial network, and the method comprises the following steps: (1) carrying out the preprocessingof images in a disclosed face image database, and inputting the images into a generator, so as to obtain a generated image; (2) inputting the real image and the generated image into a plurality of discriminators to obtain a feedback value; (3) taking the feedback values of the plurality of discriminators as confrontation loss, and performing confrontation training on the generative adversarial network by combining the perception loss and the reconstruction loss; and (4) inputting the missing face image into the trained generator to obtain a repaired face image. Aiming at the problem of repairing the shielded or damaged face image, a generative adversarial network structure with multiple discriminators is adopted, so that the problem of low authenticity of the repaired image is solved, andthe repaired image is more natural and more real.

The invention relates to a scene segmentation method and device, computer equipment and a storage medium, and relates to the technical field of machine learning. The method includes: inputting the image to be identified into the deep neural network; carrying out depth separable convolution on the image through the down-sampling module; obtaining a first feature map of which the size is smaller than that of the image; performing hole convolution on the first feature map through a hole space pyramid pooling module; obtaining a second feature map of different scales, performing depth separable convolution on the second feature maps of different scales through an up-sampling module to obtain a third feature map with the same size as the image, and classifying each pixel in the third feature map through a classification module to obtain a scene segmentation result of the image. According to the invention, the calculation amount of scene segmentation through the deep neural network can be reduced, and the accuracy of scene segmentation through the deep neural network can be ensured.

The invention discloses an image super-resolution method based on a generative adversarial network. The method comprises the following steps: obtaining a training data set and a verification data set;constructing an image super-resolution model, wherein the image super-resolution model comprises a generation network model and a discrimination network model; initializing weights of the establishedgenerative network model and the discriminant network model, initializing the network model, selecting an optimizer, and setting network training parameters; simultaneously training the generative network model and the discriminant network model by using a loss function until the generative network and the discriminant network reach Nash equilibrium; obtaining a test data set and inputting the test data set into the trained generative network model to generate a super-resolution image; and calculating a peak signal-to-noise ratio between the generated super-resolution image and a real high-resolution image, calculating an evaluation index of the image reconstruction quality of the generated image, and evaluating the reconstruction quality of the image. According to the method, the performance of reconstructing the super-resolution image by the network is improved by optimizing the network structure, and the problem of image super-resolution is solved.

The invention relates to an image super-resolution method based on a channel attention mechanism and multilayer feature fusion, and the method comprises the steps of directly extracting the original features of a low-resolution image at the beginning of a residual branch by using a single-layer convolutional layer based on deep learning; using six cascaded convolutional circulation units based ona channel attention mechanism and multi-layer feature fusion to extract accurate depth features; carrying out upsampling on the depth features through a deconvolution layer, and carrying out dimensionality reduction on the upsampled features through a single-layer convolution layer to obtain a residual error of the high-resolution image; carrying out up-sampling on the low-resolution image by using a bicubic interpolation method in a mapping branch to obtain mapping of the high-resolution image; and adding the mapping and the residual of the high-resolution image pixel by pixel to obtain a final high-resolution image. The method is reasonable in design, fully considers the difference between the feature channels, efficiently utilizes the hierarchical features, and maintains a higher operation speed while obtaining higher accuracy.

The invention discloses a meta-learning algorithm based on stepwise gradient correction of a meta-learner, and the algorithm comprises the steps: firstly, obtaining training data with noise marks anda small amount of clean unbiased metadata sets; establishing a meta-learner, namely a teacher network, on the metadata set relative to a classifier, namely a student network established on the training data set; and carrying out united updating of student network parameters and teacher network parameters by using random gradient descent; obtaining a student network parameter gradient update function through a student network gradient descent format; feeding the network parameters back to the teacher network, and updating the teacher network parameters by using metadata to obtain a corrected student network parameter gradient format; and then updating the student network parameters by using the correction format. Accordingly, the student network parameters can achieve better learning in thecorrection direction, and the over-fitting problem of noise marks is weakened. The method has the characteristics of easiness in understanding, realization, interpretability and the like of a user, and can be robustly suitable for an actual data scene containing noise marks.