4 results about "Positive contrast" patented technology

The application provides a pine dead wood recognition method, system and computer device based on aerial images, comprising: pre-processing aerial images to obtain a plurality of image blocks; inputting the plurality of image blocks into a backbone network for feature extraction to obtain a plurality of groups of multi-scale feature maps; simultaneously constructing positive contrast relationship and negative contrast relationship for each group of multi-scale feature maps to generate enhanced feature vectors; based on the enhanced feature vectors, obtaining known class confidence and unknown class exclusivity in parallel for the multi-scale candidate boxes generated for the corresponding image blocks; performing open set determination on the multi-scale candidate boxes according to the known class confidence and the unknown class exclusivity to determine each candidate box as a known class, background or unknown class; rejecting the candidate boxes determined as unknown classes and taking the remaining candidate boxes as effective candidate boxes; performing boundary box regression and non-maximum suppression on the effective candidate boxes to output the recognition result of the pine dead wood. The application does not increase unknown classes and recognizes pine dead wood to reduce false positives.

The invention provides a pine, dead and dead wood identification method and system based on an aerial image and computer equipment. The method comprises the following steps: preprocessing the aerial image to obtain a plurality of image blocks; inputting the plurality of image blocks into a backbone network for feature extraction to obtain a plurality of groups of multi-scale feature maps; simultaneously constructing a positive contrast relationship and a negative contrast relationship for each group of multi-scale feature maps to generate an enhanced feature vector; on the basis of the enhanced feature vectors, obtaining known class confidence and unknown class rejection in parallel for multi-scale candidate frames generated by corresponding image blocks; performing open set judgment on the multi-scale candidate frames according to known class confidence and unknown class rejection, so as to judge each candidate frame as a known class, a background or an unknown class; rejecting the candidate frames which are judged to be unknown classes, and taking the remaining candidate frames as effective candidate frames; and performing bounding box regression and non-maximum suppression on the effective candidate box, and outputting an identification result of the pine, withered and dead wood. According to the method, unknown classes are not added, and pine, dead and dead wood are identified to reduce false alarms.

The invention discloses a multi-degraded image progressive restoration method based on dynamic space modeling, and belongs to the field of computer vision and image processing. The method comprises the steps that (1) a progressive recovery model composed of three sub-networks is constructed, and the sub-networks comprise coding and decoding networks of different levels so as to have the global semantic feature extraction capacity and the local texture feature extraction capacity; (2) proposing a block attention unit as a basic component of the network, and performing difference reconstruction on different regions of the image by adjusting the size of an attention sub-block; (3) designing an uncertain supervised attention unit, and synchronously generating a restored image, an uncertain image and features transmitted to a next sub-network; and (4) constructing a positive contrast regularization loss function based on restored images output by the three sub-networks, and guiding the networks to carry out progressive optimization learning by taking the output of the third sub-network as a reference, taking the degraded image as a negative sample, and taking the output of the first and second sub-networks and the high-definition image as positive samples.

The invention relates to a droplet dividing and clustering method, device, medium and program, the method is applied to a droplet-based digital nucleic acid amplification quantitative analysis system, and the method comprises the following steps: reading scatter data of each droplet in a droplet fluorescence image, including position and brightness; the microdroplet fluorescence image is divided into a plurality of grids, all the grids are sequenced and numbered, all microdroplets in each grid form a sub-array, and the sub-arrays are connected to form a microdroplet array; mapping the microdroplet array into a target image with a set size by taking the index of the microdroplet in the microdroplet array as a horizontal coordinate and the brightness of the microdroplet as a vertical coordinate to perform image segmentation so as to obtain a reference microdroplet; and carrying out microdroplet division and clustering based on the reference microdroplet. According to the method, the robustness and reliability during clustering can be obviously improved, and particularly the problem that the robustness and reliability of threshold division and clustering results are poor under the conditions of different sample concentrations, different negative and positive contrast ratios, uneven light fields and the like is solved.