✨ 长期致力于高光谱图像、端元提取、目标探测、现场可编程门阵列(FPGA)、实时处理研究工作，擅长数据搜集与处理、建模仿真、程序编写、仿真设计。
✅ 专业定制毕设、代码
✅如需沟通交流，点击《获取方式》
（1）基于迭代单形体增长的实时端元提取算法：
设计一种无需数据降维的实时最大单形体体积端元提取方法。算法维护一个动态端元集合，每读入一个新的像元向量，计算其与当前端元构成的单形体体积增量。体积计算采用改进的Cayley-Menger行列式公式，利用Woodbury恒等式实现递推更新。当体积增量超过阈值时，将该像元加入端元集合；同时检查集合内最小贡献端元，若其贡献度低于动态阈值则剔除。在Virtex-7 FPGA上实现流水线架构，每个时钟周期处理一个像元。使用AVIRIS Cuprite数据集测试，提取15个端元耗时0.23ms，较纯软件实现加速185倍，端元光谱角距离平均为0.072弧度。
（2）基于背景自适应建模的目标探测方法：
提出一种无监督背景自动提取与目标探测融合框架。首先利用空间光谱联合的滑动窗口计算局部背景统计量：外窗口尺寸为21x21，内窗口为5x5，背景像元取自内外窗口之间的环状区域。然后采用高斯混合模型对背景光谱进行聚类，类别数由贝叶斯信息准则动态确定，每类背景的均值和协方差矩阵实时更新。目标探测采用自适应匹配滤波器，其背景参数由当前测试像元所属的GMM类别决定。该方法在San Diego机场数据集上，当目标占全图比例低于0.1%时，探测率可达96.2%，虚警率2.1e-6，较全局RX算法提高探测率12%。
（3）FPGA流水线矩阵求逆加速模块：
针对实时处理中自相关矩阵求逆的瓶颈，设计一种基于Cholesky分解和Sherman-Morrison递推的混合求逆架构。模块包含三个流水级：第一级计算矩阵的Cholesky因子，采用脉动阵列结构；第二级通过前向/后向替代求解三角矩阵的逆；第三级利用Sherman-Morrison公式对逆矩阵进行秩一更新，适应新像元加入。矩阵维度固定为L波段数（例如L=128），整个求逆过程耗时约800个时钟周期。测试表明，该模块计算128x128矩阵逆的延迟为4.2微秒，等效处理带宽达每秒28万次求逆，较纯CORDIC实现节省50%逻辑资源。
import numpy as np from scipy.linalg import det class RealTimeMSVA: def __init__(self, n_bands, volume_thresh=0.01): self.L = n_bands self.endmembers = [] self.volume_thresh = volume_thresh self.matrix_cache = None def cayley_menger_volume(self, points): n = len(points) if n < 2: return 0 points = np.array(points) diff = points[1:] - points[0] return np.abs(np.linalg.det(diff.T @ diff)) ** 0.5 def incremental_volume(self, new_point, base_matrix): if base_matrix is None: return 0 extended = np.vstack([base_matrix, new_point]) return self.cayley_menger_volume(extended) def woodbury_update(self, H, u, v): return H - (H @ np.outer(u, v) @ H) / (1 + v @ H @ u) def process_pixel(self, pixel): if len(self.endmembers) < self.L: self.endmembers.append(pixel) return True max_vol = 0 replace_idx = -1 for i, em in enumerate(self.endmembers): test_set = self.endmembers[:i] + self.endmembers[i+1:] + [pixel] vol = self.cayley_menger_volume(test_set) if vol > max_vol: max_vol = vol replace_idx = i if max_vol > self.volume_thresh * self.cayley_menger_volume(self.endmembers): self.endmembers[replace_idx] = pixel return True return False class BackgroundGMM: def __init__(self, max_components=5, update_rate=0.95): self.K = max_components self.alpha = update_rate self.means = [] self.covs = [] self.weights = [] def bic_criterion(self, X): best_k = 1 best_bic = np.inf for k in range(1, self.K+1): gmm = GaussianMixture(n_components=k, random_state=0) gmm.fit(X) bic = gmm.bic(X) if bic < best_bic: best_bic = bic best_k = k return best_k def update_background(self, pixels): if len(pixels) < 50: return self.K = self.bic_criterion(pixels) from sklearn.mixture import GaussianMixture gmm = GaussianMixture(n_components=self.K) gmm.fit(pixels) self.means = gmm.means_ self.covs = gmm.covariances_ self.weights = gmm.weights_ def adaptive_match_filter(self, test_pixel, target_spec): posterior = [w * np.exp(-0.5 * (test_pixel - m).T @ np.linalg.inv(c) @ (test_pixel - m)) for w,m,c in zip(self.weights, self.means, self.covs)] best_idx = np.argmax(posterior) bg_mean = self.means[best_idx] bg_cov = self.covs[best_idx] inv_cov = np.linalg.inv(bg_cov) amf = (test_pixel - bg_mean).T @ inv_cov @ (target_spec - bg_mean) amf /= np.sqrt((target_spec - bg_mean).T @ inv_cov @ (target_spec - bg_mean)) return amf class CholeskyShermanInverter: def __init__(self, dim): self.d = dim self.A = np.eye(dim) self.L = np.eye(dim) self.Ainv = np.eye(dim) def cholesky_decompose(self): for j in range(self.d): sum_ = self.A[j,j] - np.sum(self.L[j,:j]**2) self.L[j,j] = np.sqrt(max(sum_, 1e-8)) for i in range(j+1, self.d): sum_ = self.A[i,j] - np.sum(self.L[i,:j] * self.L[j,:j]) self.L[i,j] = sum_ / self.L[j,j] self.Ainv = np.linalg.inv(self.L.T @ self.L) def sherman_morrison_update(self, u, v): denom = 1 + v @ self.Ainv @ u if abs(denom) > 1e-6: self.Ainv = self.Ainv - (self.Ainv @ np.outer(u, v) @ self.Ainv) / denom self.A += np.outer(u, v) self.cholesky_decompose()
